{ "updated": "2026-10-21", "projects": [ { "slug": "claude-code", "name": "Claude Code", "tagline": "Terminal-native coding agent that plans, edits and verifies changes across a whole repo", "description": "Claude Code runs as an interactive CLI inside the repository: it maps the codebase, writes a plan, edits multiple files, runs the test suite and iterates until checks pass, then commits and opens a pull request. Sessions keep a 1M-token context warm through prompt caching, so most of the spend is re-reading context on every turn rather than generating output.", "domain": "coding", "icon": "๐Ÿง ", "featured": true, "sponsored": false, "status": "production", "verified": true, "source": { "type": "benchmark", "label": "Artificial Analysis Coding Agent Index v1.5 โ€” Claude Code vs Codex (Oct 2026)", "url": "https://artificialanalysis.ai/agents/coding-agents/comparisons/claude-code-vs-codex" }, "metrics": { "tokensPerTask": 27700000, "costPerTaskUsd": 14.19, "successRatePct": 72, "toolCallsPerTask": 266, "retryOverheadPct": 11, "wallClockMinutesPerTask": 90, "contextCompaction": "enabled" }, "usage": { "runsPerDay": 30000, "delta7dPct": 14, "basis": "community-report" }, "stack": { "frameworks": [ "custom-harness", "model-context-protocol" ], "models": [ "Claude Sonnet 5.5", "Claude Opus 5.5" ], "tools": [ "bash", "ripgrep", "git", "test-runner", "mcp-servers" ], "runtime": "Anthropic sandboxed shell (local and cloud)" }, "workflow": [ { "step": "Orient in repo", "detail": "Maps the repository, reads CLAUDE.md, lists the failing tests", "tokensPct": 6, "toolCalls": 9 }, { "step": "Plan the change", "detail": "Writes an ordered plan and picks the files to touch", "tokensPct": 7, "toolCalls": 4 }, { "step": "Edit files", "detail": "Applies multi-file edits with diff review", "tokensPct": 22, "toolCalls": 58 }, { "step": "Run tests", "detail": "Executes suites, parses failures and stack traces", "tokensPct": 18, "toolCalls": 46 }, { "step": "Debug and retry", "detail": "Reproduces each failure, patches and re-runs", "tokensPct": 31, "toolCalls": 92 }, { "step": "Verify and commit", "detail": "Reviews the final diff, writes the commit and opens a PR", "tokensPct": 16, "toolCalls": 57 } ], "deliverables": [ "Committed patch series", "Test-run report", "Pull request with rationale", "Session transcript" ], "links": { "website": "https://claude.com/claude-code", "github": "https://github.com/anthropics/claude-code", "paper": "" }, "tags": [ "coding-agent", "cli", "anthropic", "prompt-caching" ], "metricsNote": "Tokens, cost, wall time and turns per task from Artificial Analysis Coding Agent Index v1.5 (Sonnet 5.5 max)." }, { "slug": "codex-cli", "name": "Codex CLI", "tagline": "OpenAI's sandboxed terminal agent that batches tool calls to keep long tasks cheap", "description": "Codex CLI drives a sandboxed workspace from the shell, planning edits, executing commands and running tests in one batched loop instead of one action per turn. On the Artificial Analysis index it finishes coding tasks in about 15 minutes at roughly a dollar per task, with a 93% prompt-cache hit rate doing most of the cost saving.", "domain": "coding", "icon": "โŒจ๏ธ", "featured": false, "sponsored": false, "status": "production", "verified": true, "source": { "type": "benchmark", "label": "Artificial Analysis Coding Agent Index v1.5 โ€” Claude Code vs Codex (Oct 2026)", "url": "https://artificialanalysis.ai/agents/coding-agents/comparisons/claude-code-vs-codex" }, "metrics": { "tokensPerTask": 3200000, "costPerTaskUsd": 1.04, "successRatePct": 73, "toolCallsPerTask": 40, "retryOverheadPct": 9, "wallClockMinutesPerTask": 15.5, "contextCompaction": "enabled" }, "usage": { "runsPerDay": 45000, "delta7dPct": 9, "basis": "community-report" }, "stack": { "frameworks": [ "custom-harness" ], "models": [ "GPT-6.1 Sol", "GPT-6 Luna" ], "tools": [ "bash", "apply-patch", "git", "test-runner", "sandbox-exec" ], "runtime": "OpenAI Codex sandbox (seatbelt / landlock)" }, "workflow": [ { "step": "Read task and repo", "detail": "Parses the issue, inspects tree and existing tests", "tokensPct": 10, "toolCalls": 6 }, { "step": "Plan and locate", "detail": "Finds the relevant modules and drafts an edit plan", "tokensPct": 12, "toolCalls": 5 }, { "step": "Patch workspace", "detail": "Applies file patches inside the sandbox", "tokensPct": 24, "toolCalls": 12 }, { "step": "Build and test", "detail": "Runs build, linters and the test suite in batch", "tokensPct": 28, "toolCalls": 11 }, { "step": "Repair failures", "detail": "Reads failing output and issues targeted fixes", "tokensPct": 18, "toolCalls": 5 }, { "step": "Summarise", "detail": "Writes the final diff summary and command log", "tokensPct": 8, "toolCalls": 1 } ], "deliverables": [ "Patched working tree", "Test and lint output", "Unified diff summary" ], "links": { "website": "https://openai.com/codex/", "github": "https://github.com/openai/codex", "paper": "" }, "tags": [ "coding-agent", "cli", "openai", "sandbox" ], "metricsNote": "Per-task tokens, cost, time and turns from Artificial Analysis Coding Agent Index v1.5 (GPT-6.1 Sol, xhigh)." }, { "slug": "cursor-agent", "name": "Cursor Agent", "tagline": "IDE-native agent running Cursor's own Composer model for cheap multi-file edits", "description": "Cursor's agent lives inside the editor and can edit across files, run terminal commands and review its own diff, driven by Cursor's in-house Composer 2.5 model. Artificial Analysis measured $0.07 per task in standard mode and $0.44 in fast mode, roughly 10-60x cheaper than the highest-effort frontier agent variants on the same index.", "domain": "coding", "icon": "๐Ÿ–ฑ๏ธ", "featured": false, "sponsored": false, "status": "production", "verified": true, "source": { "type": "benchmark", "label": "Artificial Analysis โ€” Cursor's Composer 2.5, third on the Coding Agent Index (May 2026)", "url": "https://artificialanalysis.ai/de/articles/cursor-composer-2-5-coding-agent-index" }, "metrics": { "tokensPerTask": 350000, "costPerTaskUsd": 0.07, "successRatePct": 47, "toolCallsPerTask": 55, "retryOverheadPct": 7, "wallClockMinutesPerTask": 9.3, "contextCompaction": "enabled" }, "usage": { "runsPerDay": 500000, "delta7dPct": 6, "basis": "community-report" }, "stack": { "frameworks": [ "cursor-agent-harness" ], "models": [ "Composer 2.5", "Claude Sonnet 5.5" ], "tools": [ "terminal", "codebase-index", "apply-diff", "linter", "mcp-servers" ], "runtime": "Cursor IDE and Cursor CLI" }, "workflow": [ { "step": "Index context", "detail": "Semantic search across the codebase for relevant symbols", "tokensPct": 12, "toolCalls": 8 }, { "step": "Draft the diff", "detail": "Generates edits for the selected files", "tokensPct": 30, "toolCalls": 14 }, { "step": "Run commands", "detail": "Executes terminal commands and reads output", "tokensPct": 22, "toolCalls": 16 }, { "step": "Review own diff", "detail": "Self-reviews and fixes obvious mistakes", "tokensPct": 20, "toolCalls": 12 }, { "step": "Hand back", "detail": "Explains changes and leaves the diff for the developer", "tokensPct": 16, "toolCalls": 5 } ], "deliverables": [ "Multi-file diff", "Command log", "Inline change rationale" ], "links": { "website": "https://cursor.com", "github": "", "paper": "" }, "tags": [ "coding-agent", "ide", "composer", "cheap-inference" ], "metricsNote": "Cost per task ($0.07 standard, $0.44 fast) and wall time from Artificial Analysis; tokens derived from Cursor's published $0.50/$2.50 list price." }, { "slug": "devin", "name": "Devin", "tagline": "Autonomous cloud engineer that runs whole tickets in its own VM", "description": "Devin is Cognition's cloud software engineer: it clones the repository, plans the work, edits code, runs tests and opens pull requests inside a persistent VM, with a Fusion mode that pairs a strong lead model with a cheaper sidekick. Cognition reports FrontierCode 1.1 Extended at 68.8 for $0.60 per task after a harness rewrite that batches tool calls.", "domain": "coding", "icon": "๐Ÿ›ฐ๏ธ", "featured": false, "sponsored": false, "status": "production", "verified": true, "source": { "type": "docs", "label": "Cognition โ€” Devin is now up to 40% more cost-efficient (FrontierCode 1.1 table)", "url": "https://devin.ai/blog/more-efficient-devin" }, "metrics": { "tokensPerTask": 320000, "costPerTaskUsd": 0.6, "successRatePct": 68.8, "toolCallsPerTask": 120, "retryOverheadPct": 12, "wallClockMinutesPerTask": 55, "contextCompaction": "enabled" }, "usage": { "runsPerDay": 9500, "delta7dPct": 18, "basis": "community-report" }, "stack": { "frameworks": [ "devin-cloud-harness" ], "models": [ "Devin Fusion", "Claude Opus 5.5", "SWE-2", "GPT-6 Luna" ], "tools": [ "shell", "browser", "git", "test-runner", "slack" ], "runtime": "Cognition cloud VM snapshots" }, "workflow": [ { "step": "Accept ticket", "detail": "Reads the Jira or Slack ticket and asks clarifying questions", "tokensPct": 6, "toolCalls": 4 }, { "step": "Explore and plan", "detail": "Clones the repo, maps the code, writes a task plan", "tokensPct": 16, "toolCalls": 18 }, { "step": "Implement", "detail": "Edits files and batches format/lint/test commands", "tokensPct": 32, "toolCalls": 42 }, { "step": "Test and debug", "detail": "Runs CI locally and repairs failures", "tokensPct": 26, "toolCalls": 38 }, { "step": "Review and PR", "detail": "Self-review pass, then PR with a written summary", "tokensPct": 20, "toolCalls": 18 } ], "deliverables": [ "Pull request", "CI run log", "Implementation notes for the reviewer" ], "links": { "website": "https://devin.ai", "github": "", "paper": "" }, "tags": [ "coding-agent", "autonomous", "cloud-vm", "cognition" ], "metricsNote": "Score ($68.8) and $0.60/task from Cognition's Sept 2026 post; token count derived from the published cost at Fusion's blended list price." }, { "slug": "openhands", "name": "OpenHands", "tagline": "Open-source CodeAct agent that resolves SWE-bench issues in a sandboxed container", "description": "OpenHands (formerly OpenDevin) gives an LLM a CodeAct loop over a Docker sandbox: it writes code as actions, edits the repository, runs tests and iterates until the issue is fixed. On SWE-bench Verified it reaches around 72% resolved with inference-time scaling and a critic model, at a few dollars of tokens per instance.", "domain": "coding", "icon": "๐Ÿ› ๏ธ", "featured": false, "sponsored": false, "status": "open-source", "verified": true, "source": { "type": "community", "label": "SWE-bench deep-dive: agent performance, costs and failure taxonomy (2026)", "url": "https://raw.githubusercontent.com/MohamedAbdallah-14/Wazir/refs/heads/main/docs/research/agent-architecture/swe-bench.md" }, "metrics": { "tokensPerTask": 2400000, "costPerTaskUsd": 1.35, "successRatePct": 72, "toolCallsPerTask": 90, "retryOverheadPct": 15, "wallClockMinutesPerTask": 65, "contextCompaction": "enabled" }, "usage": { "runsPerDay": 65000, "delta7dPct": 21, "basis": "community-report" }, "stack": { "frameworks": [ "openhands", "codeact" ], "models": [ "Claude Sonnet 5.5", "GPT-6.1 Sol", "DeepSeek V4-Pro" ], "tools": [ "docker", "bash", "file-editor", "browser", "pytest" ], "runtime": "Docker sandbox with per-task workspace" }, "workflow": [ { "step": "Reproduce", "detail": "Reads the issue, reproduces the bug inside the container", "tokensPct": 12, "toolCalls": 8 }, { "step": "Localise", "detail": "Greps the repo and opens candidate files", "tokensPct": 20, "toolCalls": 22 }, { "step": "Patch", "detail": "Writes the fix as code actions in the sandbox", "tokensPct": 26, "toolCalls": 24 }, { "step": "Test loop", "detail": "Runs the failing tests, reads tracebacks, retries", "tokensPct": 28, "toolCalls": 26 }, { "step": "Critique", "detail": "A critic model checks the patch against the issue", "tokensPct": 14, "toolCalls": 10 } ], "deliverables": [ "Resolved patch", "Test evidence", "Reproduction and fix trace" ], "links": { "website": "https://openhands.dev", "github": "https://github.com/All-Hands-AI/OpenHands", "paper": "https://arxiv.org/abs/2407.16741" }, "tags": [ "coding-agent", "open-source", "swe-bench", "codeact" ], "metricsNote": "72% SWE-bench Verified from a 2026 benchmark round-up; tokens and cost per task modelled from published per-instance trajectory costs." }, { "slug": "swe-agent", "name": "SWE-agent", "tagline": "Princeton agent with a bespoke computer interface for repository tasks", "description": "SWE-agent showed that the agent-computer interface matters as much as the model: it gives the LLM a windowed file viewer, edit commands with linting feedback and a guardrailed shell. That interface work alone produced a 3.3x improvement on SWE-bench, and it remains the reference harness behind the mini-SWE-agent bash-only leaderboard.", "domain": "coding", "icon": "๐Ÿ”ง", "featured": false, "sponsored": false, "status": "open-source", "verified": false, "source": { "type": "docs", "label": "SWE-agent repository and agent-computer interface documentation", "url": "https://github.com/SWE-agent/SWE-agent" }, "metrics": { "tokensPerTask": 1200000, "costPerTaskUsd": 0.55, "successRatePct": 48, "toolCallsPerTask": 70, "retryOverheadPct": 20, "wallClockMinutesPerTask": 40, "contextCompaction": "disabled" }, "usage": { "runsPerDay": 28000, "delta7dPct": -4, "basis": "community-report" }, "stack": { "frameworks": [ "swe-agent", "mini-swe-agent" ], "models": [ "GPT-6.1 Sol", "Claude Sonnet 5.5", "DeepSeek V4.1 Flash" ], "tools": [ "bash", "file-viewer", "edit-command", "lint", "docker" ], "runtime": "Docker container per instance" }, "workflow": [ { "step": "Open issue", "detail": "Loads the problem statement and the repo layout", "tokensPct": 10, "toolCalls": 5 }, { "step": "Search files", "detail": "Uses windowed file viewer and search commands", "tokensPct": 24, "toolCalls": 20 }, { "step": "Edit with linting", "detail": "Applies edits; lint errors are fed straight back", "tokensPct": 28, "toolCalls": 18 }, { "step": "Run tests", "detail": "Executes the target tests and reads the diff output", "tokensPct": 24, "toolCalls": 18 }, { "step": "Submit patch", "detail": "Emits the final diff for evaluation", "tokensPct": 14, "toolCalls": 9 } ], "deliverables": [ "Model patch", "Lint and test feedback log", "Trajectory file" ], "links": { "website": "https://swe-agent.com", "github": "https://github.com/SWE-agent/SWE-agent", "paper": "https://arxiv.org/abs/2405.15793" }, "tags": [ "coding-agent", "open-source", "swe-bench", "agent-computer-interface" ], "metricsNote": "Estimated: per-task tokens and cost modelled from SWE-bench per-instance API costs; the 3.3x interface result is from the SWE-agent paper." }, { "slug": "aider", "name": "Aider", "tagline": "Git-native pair programmer that edits your repo from the terminal", "description": "Aider is a terminal pair programmer that maps the repository, proposes edits and commits every change to git with a generated message. Because it sends a compact repo map instead of whole files it stays among the cheapest agents per solved task: the 2026 TrustTheBench sweep put it at 67.5% on SWE-bench Verified for about $0.48 per task.", "domain": "coding", "icon": "๐Ÿ“", "featured": false, "sponsored": false, "status": "open-source", "verified": true, "source": { "type": "benchmark", "label": "TrustTheBench โ€” Autonomous Coding Agents Leaderboard 2026 (SWE-bench Verified and cost per task)", "url": "https://www.trustthebench.com/agents" }, "metrics": { "tokensPerTask": 620000, "costPerTaskUsd": 0.48, "successRatePct": 67.5, "toolCallsPerTask": 35, "retryOverheadPct": 8, "wallClockMinutesPerTask": 22, "contextCompaction": "disabled" }, "usage": { "runsPerDay": 120000, "delta7dPct": 3, "basis": "community-report" }, "stack": { "frameworks": [ "aider" ], "models": [ "Claude Sonnet 5.5", "DeepSeek V4.1 Flash", "Qwen3.8 Max" ], "tools": [ "git", "repo-map", "diff-editor", "test-command", "lint" ], "runtime": "Local terminal with git worktree" }, "workflow": [ { "step": "Build repo map", "detail": "Generates a compact map of files, symbols and sizes", "tokensPct": 14, "toolCalls": 4 }, { "step": "Locate code", "detail": "Searches for the functions named in the request", "tokensPct": 18, "toolCalls": 9 }, { "step": "Propose edit", "detail": "Emits a search/replace diff for review", "tokensPct": 26, "toolCalls": 8 }, { "step": "Apply and test", "detail": "Applies the diff and runs the test command", "tokensPct": 28, "toolCalls": 10 }, { "step": "Commit", "detail": "Writes a conventional commit for the change", "tokensPct": 14, "toolCalls": 4 } ], "deliverables": [ "Git commit", "Diff with rationale", "Test output" ], "links": { "website": "https://aider.chat", "github": "https://github.com/Aider-AI/aider", "paper": "" }, "tags": [ "coding-agent", "open-source", "git", "cli" ], "metricsNote": "Score and $0.48/task from the TrustTheBench 2026 agent sweep; token count derived from that cost at Aider's typical model mix." }, { "slug": "cline", "name": "Cline", "tagline": "Open-source VS Code agent with human approval on every tool call", "description": "Cline is the batteries-included agent for VS Code: it plans, edits files, runs terminal commands and can drive a browser, pausing for human approval on each tool call. TrustTheBench's 2026 sweep measured 66.8% on SWE-bench Verified at $0.42 per task, making it one of the cheapest ways to run an agentic edit loop on your own API key.", "domain": "coding", "icon": "๐Ÿง—", "featured": false, "sponsored": true, "status": "open-source", "verified": true, "source": { "type": "benchmark", "label": "TrustTheBench โ€” Autonomous Coding Agents Leaderboard 2026 (SWE-bench Verified and cost per task)", "url": "https://www.trustthebench.com/agents" }, "metrics": { "tokensPerTask": 540000, "costPerTaskUsd": 0.42, "successRatePct": 66.8, "toolCallsPerTask": 48, "retryOverheadPct": 10, "wallClockMinutesPerTask": 26, "contextCompaction": "enabled" }, "usage": { "runsPerDay": 95000, "delta7dPct": 25, "basis": "community-report" }, "stack": { "frameworks": [ "cline" ], "models": [ "Claude Sonnet 5.5", "GLM-4.7", "DeepSeek V4.1 Flash" ], "tools": [ "filesystem", "terminal", "browser", "mcp-servers" ], "runtime": "VS Code extension (BYOK)" }, "workflow": [ { "step": "Understand request", "detail": "Reads the task and the open files in the editor", "tokensPct": 12, "toolCalls": 5 }, { "step": "Investigate", "detail": "Searches the workspace and reads relevant files", "tokensPct": 22, "toolCalls": 14 }, { "step": "Write code", "detail": "Applies diffs after per-step approval", "tokensPct": 30, "toolCalls": 16 }, { "step": "Run and fix", "detail": "Executes commands in the integrated terminal", "tokensPct": 24, "toolCalls": 10 }, { "step": "Verify", "detail": "Checks the final diff and reports the outcome", "tokensPct": 12, "toolCalls": 3 } ], "deliverables": [ "Applied code changes", "Terminal transcript", "Approval log" ], "links": { "website": "https://cline.bot", "github": "https://github.com/cline/cline", "paper": "" }, "tags": [ "coding-agent", "open-source", "vs-code", "byok" ], "metricsNote": "Score and $0.42/task from the TrustTheBench 2026 agent sweep; token count derived from that cost across Cline's typical model mix." }, { "slug": "copilot-coding-agent", "name": "GitHub Copilot coding agent", "tagline": "Assignable cloud agent that turns GitHub issues into pull requests", "description": "The Copilot coding agent is assigned to a GitHub issue and works in its own ephemeral environment: it explores the repo, writes and runs code, then opens a pull request for review. TrustTheBench's 2026 evaluation measured 65% on SWE-bench Verified at about $0.75 of inference per task, billed through Copilot premium requests or credits.", "domain": "coding", "icon": "๐Ÿ™", "featured": false, "sponsored": false, "status": "production", "verified": true, "source": { "type": "benchmark", "label": "TrustTheBench โ€” Autonomous Coding Agents Leaderboard 2026 (SWE-bench Verified and cost per task)", "url": "https://www.trustthebench.com/agents" }, "metrics": { "tokensPerTask": 780000, "costPerTaskUsd": 0.75, "successRatePct": 65, "toolCallsPerTask": 60, "retryOverheadPct": 9, "wallClockMinutesPerTask": 30, "contextCompaction": "enabled" }, "usage": { "runsPerDay": 340000, "delta7dPct": 12, "basis": "community-report" }, "stack": { "frameworks": [ "copilot-agent-harness" ], "models": [ "GPT-6.1 Sol", "Claude Sonnet 5.5" ], "tools": [ "github-actions", "code-search", "bash", "test-runner", "codespaces" ], "runtime": "GitHub-hosted ephemeral runner" }, "workflow": [ { "step": "Read the issue", "detail": "Parses the issue, linked files and repository conventions", "tokensPct": 8, "toolCalls": 4 }, { "step": "Explore repository", "detail": "Uses code search and file reads to locate the change", "tokensPct": 20, "toolCalls": 14 }, { "step": "Implement", "detail": "Edits files and stages the change on a branch", "tokensPct": 28, "toolCalls": 18 }, { "step": "Test in runner", "detail": "Runs the build and tests inside the ephemeral environment", "tokensPct": 26, "toolCalls": 16 }, { "step": "Open PR", "detail": "Writes the PR body and requests review", "tokensPct": 18, "toolCalls": 8 } ], "deliverables": [ "Pull request on a branch", "CI results", "Agent session log" ], "links": { "website": "https://github.com/features/copilot", "github": "https://github.com/features/copilot", "paper": "" }, "tags": [ "coding-agent", "github", "cloud-agent", "pull-request" ], "metricsNote": "Score and $0.75/task from the TrustTheBench 2026 agent sweep; tokens derived from that cost at Copilot's model mix." }, { "slug": "jules", "name": "Jules", "tagline": "Google's async coding agent that clones repos and returns diffs without a laptop", "description": "Jules takes an assigned task, spins up a cloud VM with a clone of the repository, plans the change, edits and tests the code, then hands back a patch or pull request asynchronously. It is built for background work: you queue a task, close the tab, and review a diff later against a Google account quota rather than a per-token meter.", "domain": "coding", "icon": "๐Ÿš€", "featured": false, "sponsored": false, "status": "production", "verified": false, "source": { "type": "docs", "label": "Jules โ€” Google's asynchronous coding agent (product and pricing docs)", "url": "https://jules.google" }, "metrics": { "tokensPerTask": 1100000, "costPerTaskUsd": 0.95, "successRatePct": 60, "toolCallsPerTask": 75, "retryOverheadPct": 11, "wallClockMinutesPerTask": 45, "contextCompaction": "unknown" }, "usage": { "runsPerDay": 42000, "delta7dPct": -6, "basis": "estimate" }, "stack": { "frameworks": [ "jules-harness" ], "models": [ "Gemini 3.1 Pro", "Gemini 3.8 Flash" ], "tools": [ "git", "bash", "test-runner", "web-search" ], "runtime": "Google Cloud VM per task" }, "workflow": [ { "step": "Queue task", "detail": "Accepts the prompt, repo and branch to work on", "tokensPct": 6, "toolCalls": 2 }, { "step": "Clone and inspect", "detail": "Clones the repo and reads the relevant modules", "tokensPct": 20, "toolCalls": 16 }, { "step": "Plan", "detail": "Produces a numbered plan before editing", "tokensPct": 10, "toolCalls": 3 }, { "step": "Implement and test", "detail": "Edits files, runs the suite and iterates", "tokensPct": 44, "toolCalls": 40 }, { "step": "Publish diff", "detail": "Creates a branch and pull request for review", "tokensPct": 20, "toolCalls": 14 } ], "deliverables": [ "Branch with commits", "Pull request", "Task activity log" ], "links": { "website": "https://jules.google", "github": "", "paper": "" }, "tags": [ "coding-agent", "google", "async", "cloud-vm" ], "metricsNote": "Estimated: tokens and cost modelled from Google's plan quotas and comparable cloud-agent task profiles; no public per-task token count." }, { "slug": "ai-scientist-v2", "name": "AI Scientist v2", "tagline": "Autonomous system that runs the full research loop and writes a paper", "description": "The AI Scientist generates hypotheses from a topic, designs and runs experiments in a sandbox, analyses results, writes a full manuscript in LaTeX and reviews its own output. Sakana's system became the first autonomous pipeline with a paper published in Nature, and independent evaluation puts a complete manuscript at $6-15 and about 3.5 hours of human supervision.", "domain": "research", "icon": "๐Ÿ”ฌ", "featured": true, "sponsored": false, "status": "research", "verified": true, "source": { "type": "paper", "label": "The AI Scientist v2 โ€” fully automated open-ended scientific discovery (paper)", "url": "https://arxiv.org/abs/2504.08066" }, "metrics": { "tokensPerTask": 6200000, "costPerTaskUsd": 15, "successRatePct": 42, "toolCallsPerTask": 180, "retryOverheadPct": 18, "wallClockMinutesPerTask": 210, "contextCompaction": "enabled" }, "usage": { "runsPerDay": 1900, "delta7dPct": 30, "basis": "estimate" }, "stack": { "frameworks": [ "ai-scientist-v2", "aider" ], "models": [ "GPT-6.1 Sol", "Claude Sonnet 5.5", "DeepSeek V4-Pro" ], "tools": [ "python", "latex", "experiment-runner", "web-search", "figure-generator" ], "runtime": "Local GPU box or cloud VM with sandboxed experiments" }, "workflow": [ { "step": "Generate ideas", "detail": "Surveys the topic and proposes candidate hypotheses", "tokensPct": 8, "toolCalls": 12 }, { "step": "Design experiment", "detail": "Writes the experiment plan and code skeleton", "tokensPct": 12, "toolCalls": 16 }, { "step": "Run experiments", "detail": "Executes training runs, monitors logs, fixes crashes", "tokensPct": 34, "toolCalls": 74 }, { "step": "Analyse results", "detail": "Parses metrics, makes plots, decides what worked", "tokensPct": 18, "toolCalls": 28 }, { "step": "Write manuscript", "detail": "Drafts the LaTeX paper with figures and citations", "tokensPct": 20, "toolCalls": 32 }, { "step": "Self-review", "detail": "Runs an automated reviewer and revises", "tokensPct": 8, "toolCalls": 18 } ], "deliverables": [ "LaTeX manuscript", "Experiment code and logs", "Figures and result tables", "Automated review notes" ], "links": { "website": "https://sakana.ai/ai-scientist/", "github": "https://github.com/SakanaAI/AI-Scientist-v2", "paper": "https://arxiv.org/abs/2504.08066" }, "tags": [ "autonomous-research", "scientific-discovery", "paper-writing" ], "metricsNote": "Cost per paper ($6-15) from independent evaluation and Sakana's reporting; token count derived from that cost at the pipeline's model mix." }, { "slug": "futurehouse-robin", "name": "FutureHouse Robin", "tagline": "Multi-agent lab assistant that proposes and validates chemistry hypotheses", "description": "Robin orchestrates specialised agents (literature search, hypothesis generation, experiment design and data analysis) over a chemistry knowledge base, then proposes candidate treatments and checks them against assay data. In its published dry-age-related macular degeneration case it identified and validated ripasudil as a candidate in a few days of agent time.", "domain": "research", "icon": "๐Ÿงช", "featured": false, "sponsored": false, "status": "production", "verified": false, "source": { "type": "paper", "label": "FutureHouse โ€” Robin: a multi-agent system for automating scientific discovery", "url": "https://www.futurehouse.org/research/robin" }, "metrics": { "tokensPerTask": 4500000, "costPerTaskUsd": 12, "successRatePct": 55, "toolCallsPerTask": 140, "retryOverheadPct": 22, "wallClockMinutesPerTask": 180, "contextCompaction": "enabled" }, "usage": { "runsPerDay": 320, "delta7dPct": 8, "basis": "estimate" }, "stack": { "frameworks": [ "robin", "langgraph" ], "models": [ "Claude Opus 5.5", "GPT-6.1 Sol" ], "tools": [ "literature-search", "chem-database", "python", "lab-automation-api" ], "runtime": "FutureHouse cloud platform" }, "workflow": [ { "step": "Frame question", "detail": "Turns the biological question into searchable sub-questions", "tokensPct": 8, "toolCalls": 8 }, { "step": "Literature sweep", "detail": "Queries papers and internal assays, extracts findings", "tokensPct": 26, "toolCalls": 44 }, { "step": "Generate hypotheses", "detail": "Proposes candidate mechanisms and molecules", "tokensPct": 18, "toolCalls": 20 }, { "step": "Rank and design", "detail": "Scores candidates and designs validation experiments", "tokensPct": 24, "toolCalls": 30 }, { "step": "Analyse data", "detail": "Reads assay output and revises the hypothesis", "tokensPct": 24, "toolCalls": 38 } ], "deliverables": [ "Ranked candidate list", "Evidence report with citations", "Proposed validation experiments" ], "links": { "website": "https://www.futurehouse.org", "github": "", "paper": "https://www.futurehouse.org/research/robin" }, "tags": [ "autonomous-research", "chemistry", "multi-agent", "drug-discovery" ], "metricsNote": "Estimated: no public per-task token or dollar figures; modelled from comparable multi-agent literature-plus-analysis workloads." }, { "slug": "gpt-researcher", "name": "GPT Researcher", "tagline": "Open-source deep-research agent that writes cited reports from 20+ sources", "description": "GPT Researcher plans a research question, scrapes and filters 20+ sources in parallel, then aggregates them into a cited report with a planner and a publisher agent. The project documents roughly $0.01 of model spend per research run, which makes it the cheapest credible deep-research loop to self-host.", "domain": "research", "icon": "๐Ÿ“š", "featured": false, "sponsored": false, "status": "open-source", "verified": true, "source": { "type": "docs", "label": "GPT Researcher FAQ โ€” cost per research run", "url": "https://docs.gptr.dev/docs/faq" }, "metrics": { "tokensPerTask": 34000, "costPerTaskUsd": 0.01, "successRatePct": 88, "toolCallsPerTask": 22, "retryOverheadPct": 6, "wallClockMinutesPerTask": 4, "contextCompaction": "disabled" }, "usage": { "runsPerDay": 58000, "delta7dPct": 17, "basis": "community-report" }, "stack": { "frameworks": [ "gpt-researcher", "langchain" ], "models": [ "GPT-6 Luna", "DeepSeek V4.1 Flash" ], "tools": [ "web-scraper", "tavily-search", "embeddings", "pdf-reader" ], "runtime": "Python service, Docker image" }, "workflow": [ { "step": "Plan queries", "detail": "Breaks the question into search sub-queries", "tokensPct": 10, "toolCalls": 3 }, { "step": "Scrape sources", "detail": "Fetches and filters 20+ pages concurrently", "tokensPct": 26, "toolCalls": 12 }, { "step": "Summarise", "detail": "Condenses each source into evidence notes", "tokensPct": 32, "toolCalls": 3 }, { "step": "Write report", "detail": "Aggregates notes into a structured cited report", "tokensPct": 24, "toolCalls": 2 }, { "step": "Check citations", "detail": "Verifies links and drops unsupported claims", "tokensPct": 8, "toolCalls": 2 } ], "deliverables": [ "Markdown research report", "Source list with URLs", "Per-source evidence notes" ], "links": { "website": "https://gptr.dev", "github": "https://github.com/assafelovic/gpt-researcher", "paper": "https://arxiv.org/abs/2403.06634" }, "tags": [ "deep-research", "open-source", "rag", "citations" ], "metricsNote": "Docs state about $0.01 per research run; token count derived from that cost at current cheap-model list prices." }, { "slug": "coscientist", "name": "Coscientist", "tagline": "LLM agent that designs, codes and runs real chemistry experiments", "description": "Coscientist combines web search, code execution and robotic lab APIs so an LLM can plan a synthesis, write the control code and actually run it on automated hardware. In its Nature paper it optimised Suzuki and Sonogashira couplings and discovered a new reaction pathway, working end-to-end from a natural-language brief.", "domain": "research", "icon": "โš—๏ธ", "featured": false, "sponsored": false, "status": "research", "verified": false, "source": { "type": "paper", "label": "Nature โ€” Autonomous chemical research with large language models (Coscientist)", "url": "https://www.nature.com/articles/s41586-023-06792-0" }, "metrics": { "tokensPerTask": 2100000, "costPerTaskUsd": 6.5, "successRatePct": 65, "toolCallsPerTask": 95, "retryOverheadPct": 14, "wallClockMinutesPerTask": 120, "contextCompaction": "unknown" }, "usage": { "runsPerDay": 85, "delta7dPct": 2, "basis": "estimate" }, "stack": { "frameworks": [ "custom-harness" ], "models": [ "GPT-6.1 Sol", "Claude Opus 5.5" ], "tools": [ "web-search", "python", "cloud-lab-api", "liquid-handler", "chromatography-reader" ], "runtime": "Cloud lab with robotic synthesis hardware" }, "workflow": [ { "step": "Parse objective", "detail": "Turns the chemistry goal into a plan and constraints", "tokensPct": 8, "toolCalls": 4 }, { "step": "Search protocols", "detail": "Reads literature and vendor docs for procedures", "tokensPct": 20, "toolCalls": 18 }, { "step": "Write control code", "detail": "Generates and tests the Python that drives the robot", "tokensPct": 26, "toolCalls": 22 }, { "step": "Execute runs", "detail": "Runs synthesis cycles and reads instrument output", "tokensPct": 30, "toolCalls": 34 }, { "step": "Analyse and iterate", "detail": "Interprets spectra and adjusts the next run", "tokensPct": 16, "toolCalls": 17 } ], "deliverables": [ "Synthesis protocol", "Instrument control code", "Reaction yield analysis" ], "links": { "website": "https://www.nature.com/articles/s41586-023-06792-0", "github": "https://github.com/gomes-lab/coscientist", "paper": "https://www.nature.com/articles/s41586-023-06792-0" }, "tags": [ "autonomous-research", "chemistry", "lab-automation" ], "metricsNote": "Estimated: the paper reports capability, not token spend; per-task tokens and cost modelled from a long search-plus-code-execution loop." }, { "slug": "elicit", "name": "Elicit", "tagline": "Literature-review agent that extracts structured data from papers at scale", "description": "Elicit runs a systematic-review loop over the academic corpus: it searches papers, screens them against inclusion criteria, extracts chosen columns from each PDF and returns a review table with citations. Researchers use it to turn hundreds of papers into a comparable dataset without reading every abstract by hand.", "domain": "research", "icon": "๐Ÿ“„", "featured": false, "sponsored": false, "status": "production", "verified": false, "source": { "type": "docs", "label": "Elicit โ€” AI research assistant for systematic reviews (product docs)", "url": "https://elicit.com" }, "metrics": { "tokensPerTask": 180000, "costPerTaskUsd": 0.35, "successRatePct": 82, "toolCallsPerTask": 30, "retryOverheadPct": 5, "wallClockMinutesPerTask": 12, "contextCompaction": "enabled" }, "usage": { "runsPerDay": 60000, "delta7dPct": 11, "basis": "community-report" }, "stack": { "frameworks": [ "custom-harness" ], "models": [ "Claude Sonnet 5.5", "Gemini 3.8 Flash" ], "tools": [ "semantic-scholar", "pdf-parser", "table-extractor" ], "runtime": "Elicit cloud (web app)" }, "workflow": [ { "step": "Search corpus", "detail": "Queries 125M+ papers and ranks candidates", "tokensPct": 12, "toolCalls": 4 }, { "step": "Screen abstracts", "detail": "Applies inclusion and exclusion criteria", "tokensPct": 26, "toolCalls": 8 }, { "step": "Extract columns", "detail": "Pulls study design, n, outcome and effect size per paper", "tokensPct": 40, "toolCalls": 12 }, { "step": "Build table", "detail": "Normalises and deduplicates extracted values", "tokensPct": 14, "toolCalls": 4 }, { "step": "Cite", "detail": "Attaches page-level citations to every cell", "tokensPct": 8, "toolCalls": 2 } ], "deliverables": [ "Structured review table", "Screened paper list", "Cell-level citations" ], "links": { "website": "https://elicit.com", "github": "", "paper": "" }, "tags": [ "literature-review", "extraction", "systematic-review" ], "metricsNote": "Estimated: Elicit publishes seat pricing, not token spend; per-task tokens modelled from a 40-paper extraction run." }, { "slug": "agentforce", "name": "Salesforce Agentforce", "tagline": "CRM-native service agent billed per action, not per seat", "description": "Agentforce deploys service and sales agents inside Salesforce data, grounding answers in CRM records and executing actions such as case updates, refunds and order changes. Consumption is metered in Flex Credits: a standard action costs 20 credits (about $0.10), with a 10K-token threshold per action before extra credits apply.", "domain": "enterprise", "icon": "โ˜๏ธ", "featured": true, "sponsored": false, "status": "production", "verified": true, "source": { "type": "community", "label": "Agentforce Flex Credits cost model 2026 โ€” 20 credits โ‰ˆ $0.10 per action, 10K-token action threshold", "url": "https://www.jitendrazaa.com/blog/salesforce/salesforce-agentforce-credits-cost-model-complete-guide-2026/" }, "metrics": { "tokensPerTask": 42000, "costPerTaskUsd": 0.42, "successRatePct": 78, "toolCallsPerTask": 4, "retryOverheadPct": 6, "wallClockMinutesPerTask": 8, "contextCompaction": "disabled" }, "usage": { "runsPerDay": 90000, "delta7dPct": 19, "basis": "community-report" }, "stack": { "frameworks": [ "agentforce-atlas-reasoning-engine" ], "models": [ "Einstein Trust Layer model mix", "Claude Sonnet 5.5" ], "tools": [ "salesforce-crm", "flow", "apex", "knowledge-base", "data-cloud" ], "runtime": "Salesforce Hyperforce (per-org)" }, "workflow": [ { "step": "Ground context", "detail": "Pulls the customer record and entitlements from CRM", "tokensPct": 18, "toolCalls": 1 }, { "step": "Plan action", "detail": "Chooses a flow, prompt template or Apex action", "tokensPct": 14, "toolCalls": 1 }, { "step": "Execute action", "detail": "Runs the CRM write or external API call", "tokensPct": 30, "toolCalls": 1 }, { "step": "Compose reply", "detail": "Writes the channel-appropriate answer with citations", "tokensPct": 26, "toolCalls": 1 }, { "step": "Escalate or close", "detail": "Hands off to a human or closes the case", "tokensPct": 12, "toolCalls": 0 } ], "deliverables": [ "Case resolution", "CRM action log", "Grounded answer with record citations" ], "links": { "website": "https://www.salesforce.com/agentforce/", "github": "", "paper": "" }, "tags": [ "enterprise", "crm", "customer-service", "consumption-pricing" ], "metricsNote": "Cost per conversation derived from published 20-credit ($0.10) actions; 10K-token action threshold and credits-per-action are public." }, { "slug": "copilot-studio", "name": "Microsoft Copilot Studio", "tagline": "Low-code agent builder billed in Copilot Credits per answer and action", "description": "Copilot Studio lets teams assemble agents over Microsoft 365 knowledge, Dataverse and Power Automate flows, with generative answers and agent actions metered in Copilot Credits. The official rate card prices a generative answer at 2 credits, an agent action at 5 credits and tenant-graph grounding at 10 credits, which makes per-interaction cost predictable.", "domain": "enterprise", "icon": "๐ŸชŸ", "featured": false, "sponsored": false, "status": "production", "verified": true, "source": { "type": "docs", "label": "Microsoft Learn โ€” Copilot Studio billing rates and Copilot Credits", "url": "https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-messages-management" }, "metrics": { "tokensPerTask": 12500, "costPerTaskUsd": 0.05, "successRatePct": 74, "toolCallsPerTask": 3, "retryOverheadPct": 8, "wallClockMinutesPerTask": 6, "contextCompaction": "disabled" }, "usage": { "runsPerDay": 120000, "delta7dPct": 15, "basis": "community-report" }, "stack": { "frameworks": [ "copilot-studio" ], "models": [ "GPT-6.1 Sol", "Claude Sonnet 5.5" ], "tools": [ "dataverse", "power-automate", "microsoft-graph", "sharepoint", "connectors" ], "runtime": "Microsoft 365 tenant" }, "workflow": [ { "step": "Retrieve knowledge", "detail": "Grounds the answer in SharePoint, Dataverse and Graph", "tokensPct": 22, "toolCalls": 1 }, { "step": "Classify intent", "detail": "Decides answer vs action vs handoff", "tokensPct": 12, "toolCalls": 0 }, { "step": "Generate answer", "detail": "Writes the response with citations", "tokensPct": 34, "toolCalls": 1 }, { "step": "Run action", "detail": "Calls a Power Automate flow or connector", "tokensPct": 24, "toolCalls": 1 }, { "step": "Log and route", "detail": "Writes the transcript and routes escalations", "tokensPct": 8, "toolCalls": 0 } ], "deliverables": [ "Grounded answer", "Automation run", "Conversation transcript" ], "links": { "website": "https://www.microsoft.com/en-us/microsoft-copilot-studio", "github": "", "paper": "" }, "tags": [ "enterprise", "low-code", "microsoft", "credits" ], "metricsNote": "Credit rates from Microsoft Learn (2 credits/generative answer, 5/action); tokens per interaction estimated from a typical 3-turn conversation." }, { "slug": "sierra", "name": "Sierra", "tagline": "Outcome-priced customer agents that own the whole resolution", "description": "Sierra builds customer-facing agents across chat, voice and email that connect to CRMs, order systems and data warehouses and complete actions rather than just answer questions. Pricing is outcome-based and negotiated per resolved conversation, so the interesting metric is cost per resolution instead of cost per token.", "domain": "enterprise", "icon": "๐Ÿ’ฌ", "featured": false, "sponsored": false, "status": "production", "verified": false, "source": { "type": "community", "label": "The AI Agent Index โ€” Sierra review, outcome-based pricing (Oct 2026)", "url": "https://theaiagentindex.com/agents/sierra" }, "metrics": { "tokensPerTask": 38000, "costPerTaskUsd": 1.85, "successRatePct": 81, "toolCallsPerTask": 5, "retryOverheadPct": 9, "wallClockMinutesPerTask": 7, "contextCompaction": "enabled" }, "usage": { "runsPerDay": 180000, "delta7dPct": 22, "basis": "estimate" }, "stack": { "frameworks": [ "sierra-agent-os" ], "models": [ "Claude Sonnet 5.5", "GPT-6.1 Sol" ], "tools": [ "crm", "order-management", "subscription-billing", "knowledge-base", "voice-stack" ], "runtime": "Sierra managed cloud (multi-channel)" }, "workflow": [ { "step": "Identify customer", "detail": "Authenticates and loads order and subscription context", "tokensPct": 16, "toolCalls": 1 }, { "step": "Diagnose intent", "detail": "Classifies the request and checks policy", "tokensPct": 18, "toolCalls": 1 }, { "step": "Take action", "detail": "Issues refund, changes plan or updates address", "tokensPct": 28, "toolCalls": 1 }, { "step": "Confirm", "detail": "Reads back the change and sets expectations", "tokensPct": 22, "toolCalls": 1 }, { "step": "Follow up", "detail": "Sends confirmation and logs the resolution", "tokensPct": 16, "toolCalls": 1 } ], "deliverables": [ "Resolved conversation", "Action record in the system of record", "Post-contact summary" ], "links": { "website": "https://sierra.ai", "github": "", "paper": "" }, "tags": [ "enterprise", "customer-service", "voice", "outcome-pricing" ], "metricsNote": "Estimated: Sierra negotiates per-resolution pricing privately; $1.85 per resolution modelled from reported enterprise outcome deals." }, { "slug": "decagon", "name": "Decagon", "tagline": "Support agent that resolves tickets across email, chat and voice", "description": "Decagon ingests help-centre content, past tickets and product APIs, then resolves support conversations end-to-end, escalating only when policy or confidence requires it. The company argues for outcome-based pricing, charging per resolution rather than per seat, so token cost is buried inside the resolution price.", "domain": "enterprise", "icon": "๐ŸŽง", "featured": false, "sponsored": false, "status": "production", "verified": false, "source": { "type": "docs", "label": "Decagon โ€” Pricing the AI agent economy (outcome-based pricing)", "url": "https://decagon.ai/blog/pricing-ai-agents" }, "metrics": { "tokensPerTask": 32000, "costPerTaskUsd": 1.2, "successRatePct": 79, "toolCallsPerTask": 5, "retryOverheadPct": 10, "wallClockMinutesPerTask": 7, "contextCompaction": "enabled" }, "usage": { "runsPerDay": 95000, "delta7dPct": 27, "basis": "estimate" }, "stack": { "frameworks": [ "decagon-platform" ], "models": [ "Claude Sonnet 5.5", "GPT-6.1 Sol" ], "tools": [ "zendesk", "salesforce", "stripe", "knowledge-base", "api-connectors" ], "runtime": "Decagon managed cloud" }, "workflow": [ { "step": "Triage", "detail": "Classifies the ticket and pulls account history", "tokensPct": 18, "toolCalls": 1 }, { "step": "Retrieve policy", "detail": "Searches help centre and internal policy docs", "tokensPct": 22, "toolCalls": 1 }, { "step": "Act", "detail": "Performs the refund, cancellation or account fix", "tokensPct": 28, "toolCalls": 1 }, { "step": "Reply", "detail": "Drafts the customer-facing answer", "tokensPct": 20, "toolCalls": 1 }, { "step": "QA and log", "detail": "Scores the resolution and writes the summary", "tokensPct": 12, "toolCalls": 1 } ], "deliverables": [ "Resolved ticket", "Action audit trail", "CSAT-ready summary" ], "links": { "website": "https://decagon.ai", "github": "", "paper": "" }, "tags": [ "enterprise", "customer-support", "outcome-pricing" ], "metricsNote": "Estimated: Decagon does not publish a rate card; $1.20 per resolution modelled from reported outcome-based contracts." }, { "slug": "harvey", "name": "Harvey", "tagline": "Legal agent that drafts, reviews and cites across matter files", "description": "Harvey works inside law-firm workflows on diligence, contract review, litigation research and drafting, grounded in the firm's own document set and playbooks. It is sold per seat with usage limits, and firms typically measure it by hours saved per matter rather than tokens per query.", "domain": "enterprise", "icon": "โš–๏ธ", "featured": false, "sponsored": false, "status": "production", "verified": false, "source": { "type": "docs", "label": "Harvey โ€” professional-grade AI for legal work (product and platform docs)", "url": "https://www.harvey.ai" }, "metrics": { "tokensPerTask": 260000, "costPerTaskUsd": 4.8, "successRatePct": 76, "toolCallsPerTask": 18, "retryOverheadPct": 12, "wallClockMinutesPerTask": 25, "contextCompaction": "enabled" }, "usage": { "runsPerDay": 32000, "delta7dPct": 13, "basis": "estimate" }, "stack": { "frameworks": [ "harvey-platform" ], "models": [ "Claude Opus 5.5", "GPT-6.1 Sol" ], "tools": [ "document-store", "citation-engine", "playbook-library", "dms-connectors" ], "runtime": "Harvey cloud (firm tenancy)" }, "workflow": [ { "step": "Load matter", "detail": "Indexes the data room, contracts and prior work product", "tokensPct": 14, "toolCalls": 2 }, { "step": "Research", "detail": "Searches precedent and authority with citations", "tokensPct": 24, "toolCalls": 5 }, { "step": "Analyse", "detail": "Extracts obligations, risks and deviations from playbook", "tokensPct": 28, "toolCalls": 5 }, { "step": "Draft", "detail": "Produces the memo, redline or clause set", "tokensPct": 26, "toolCalls": 4 }, { "step": "Verify citations", "detail": "Checks every citation against the source document", "tokensPct": 8, "toolCalls": 2 } ], "deliverables": [ "Diligence memo", "Redlined contract", "Citation-checked research note" ], "links": { "website": "https://www.harvey.ai", "github": "", "paper": "" }, "tags": [ "enterprise", "legal", "diligence", "citations" ], "metricsNote": "Estimated: Harvey sells seats, not tokens; per-matter usage modelled from a 200-document diligence review." }, { "slug": "jasper", "name": "Jasper", "tagline": "Brand-trained marketing agent that ships campaigns end to end", "description": "Jasper runs campaign work on top of a brand voice and style guide: it drafts multi-channel copy, generates variants, checks claims against brand rules and exports to the CMS or ad platform. Teams buy seats plus credits, so the cost per deliverable depends on how much of the campaign the agent assembles.", "domain": "marketing", "icon": "๐ŸŽจ", "featured": false, "sponsored": false, "status": "production", "verified": false, "source": { "type": "docs", "label": "Jasper โ€” AI marketing platform plans and pricing", "url": "https://www.jasper.ai/pricing" }, "metrics": { "tokensPerTask": 65000, "costPerTaskUsd": 0.85, "successRatePct": 80, "toolCallsPerTask": 6, "retryOverheadPct": 7, "wallClockMinutesPerTask": 5, "contextCompaction": "disabled" }, "usage": { "runsPerDay": 46000, "delta7dPct": -3, "basis": "estimate" }, "stack": { "frameworks": [ "jasper-agents" ], "models": [ "GPT-6.1 Sol", "Claude Sonnet 5.5" ], "tools": [ "brand-voice", "cms-connector", "ad-platforms", "image-generator" ], "runtime": "Jasper SaaS workspace" }, "workflow": [ { "step": "Load brand context", "detail": "Applies voice, tone and banned-claims rules", "tokensPct": 14, "toolCalls": 1 }, { "step": "Plan campaign", "detail": "Breaks the brief into channel deliverables", "tokensPct": 14, "toolCalls": 1 }, { "step": "Draft assets", "detail": "Writes copy variants for each channel", "tokensPct": 34, "toolCalls": 2 }, { "step": "Check claims", "detail": "Validates against compliance rules", "tokensPct": 20, "toolCalls": 1 }, { "step": "Export", "detail": "Pushes approved assets to the CMS or ad account", "tokensPct": 18, "toolCalls": 1 } ], "deliverables": [ "Campaign copy set", "Channel-specific variants", "Compliance-checked asset list" ], "links": { "website": "https://www.jasper.ai", "github": "", "paper": "" }, "tags": [ "marketing", "brand-voice", "campaigns" ], "metricsNote": "Estimated: Jasper bills seats plus credits, not tokens; per-campaign usage modelled from a 12-asset launch." }, { "slug": "copy-ai", "name": "Copy.ai", "tagline": "GTM agent that researches accounts and drafts outbound sequences", "description": "Copy.ai's GTM agents chain research, enrichment and writing: they pull account and contact data, score fit, draft personalised sequences and push them into the CRM or sales-engagement tool. It is positioned as a revenue workflow agent rather than a copy generator.", "domain": "marketing", "icon": "โœ๏ธ", "featured": false, "sponsored": false, "status": "production", "verified": false, "source": { "type": "docs", "label": "Copy.ai โ€” GTM AI platform agents and pricing", "url": "https://www.copy.ai/pricing" }, "metrics": { "tokensPerTask": 90000, "costPerTaskUsd": 0.65, "successRatePct": 77, "toolCallsPerTask": 9, "retryOverheadPct": 11, "wallClockMinutesPerTask": 8, "contextCompaction": "disabled" }, "usage": { "runsPerDay": 38000, "delta7dPct": -7, "basis": "estimate" }, "stack": { "frameworks": [ "copy-ai-workflows" ], "models": [ "GPT-6.1 Sol", "GPT-6 Luna" ], "tools": [ "crm", "enrichment-apis", "linkedin-data", "email-sequencer" ], "runtime": "Copy.ai SaaS workspace" }, "workflow": [ { "step": "Research account", "detail": "Pulls firmographics, news and tech stack", "tokensPct": 22, "toolCalls": 2 }, { "step": "Score fit", "detail": "Ranks against the ideal customer profile", "tokensPct": 16, "toolCalls": 1 }, { "step": "Find contacts", "detail": "Enriches and validates decision makers", "tokensPct": 20, "toolCalls": 2 }, { "step": "Draft sequence", "detail": "Writes 4-6 personalised touches", "tokensPct": 30, "toolCalls": 3 }, { "step": "Sync CRM", "detail": "Pushes the sequence and updates records", "tokensPct": 12, "toolCalls": 1 } ], "deliverables": [ "Account research brief", "Personalised outbound sequence", "CRM activity record" ], "links": { "website": "https://www.copy.ai", "github": "", "paper": "" }, "tags": [ "marketing", "gtm", "outbound", "enrichment" ], "metricsNote": "Estimated: seat-plus-credit pricing; per-sequence usage modelled from a 40-account campaign." }, { "slug": "albert", "name": "Albert.ai", "tagline": "Autonomous ad-ops agent that allocates spend and iterates creative", "description": "Albert connects to ad platforms and analytics, then runs the optimisation loop itself: it segments audiences, allocates budget, generates creative variations and reallocates spend based on measured lift. The platform reports on incremental return rather than clicks, and is priced as a platform fee plus media management.", "domain": "marketing", "icon": "๐Ÿ“ˆ", "featured": false, "sponsored": false, "status": "production", "verified": false, "source": { "type": "docs", "label": "Albert.ai โ€” autonomous digital marketing platform", "url": "https://albert.ai" }, "metrics": { "tokensPerTask": 140000, "costPerTaskUsd": 1.1, "successRatePct": 71, "toolCallsPerTask": 14, "retryOverheadPct": 13, "wallClockMinutesPerTask": 15, "contextCompaction": "enabled" }, "usage": { "runsPerDay": 15000, "delta7dPct": 5, "basis": "estimate" }, "stack": { "frameworks": [ "albert-autonomous-platform" ], "models": [ "GPT-6.1 Sol", "Gemini 3.1 Pro" ], "tools": [ "ads-manager", "analytics", "dsp", "creative-generator", "cdn" ], "runtime": "Albert managed cloud" }, "workflow": [ { "step": "Ingest performance", "detail": "Pulls spend, conversion and audience data from ad platforms", "tokensPct": 18, "toolCalls": 2 }, { "step": "Diagnose", "detail": "Finds underperforming segments and creative fatigue", "tokensPct": 22, "toolCalls": 2 }, { "step": "Decide budget", "detail": "Computes reallocation across channels and audiences", "tokensPct": 20, "toolCalls": 2 }, { "step": "Generate creative", "detail": "Produces and variants ad copy and assets", "tokensPct": 26, "toolCalls": 4 }, { "step": "Launch and measure", "detail": "Pushes changes and measures incremental lift", "tokensPct": 14, "toolCalls": 4 } ], "deliverables": [ "Budget reallocation plan", "Creative variants", "Incrementality report" ], "links": { "website": "https://albert.ai", "github": "", "paper": "" }, "tags": [ "marketing", "ad-ops", "budget-allocation" ], "metricsNote": "Estimated: Albert charges a platform fee plus media spend; per-cycle token cost modelled from a weekly optimisation run." }, { "slug": "synthesia", "name": "Synthesia", "tagline": "Video agent that turns a document into a presenter-led video", "description": "Synthesia's agent takes a script, deck or document and produces a presenter-led video: it plans scenes, writes the script, generates the avatar performance, adds captions and localises into 140+ languages. Enterprise teams use it for training and internal comms at a fraction of studio cost.", "domain": "marketing", "icon": "๐ŸŽฌ", "featured": true, "sponsored": false, "status": "production", "verified": false, "source": { "type": "docs", "label": "Synthesia โ€” AI video platform plans and pricing", "url": "https://www.synthesia.io/pricing" }, "metrics": { "tokensPerTask": 220000, "costPerTaskUsd": 3.2, "successRatePct": 84, "toolCallsPerTask": 8, "retryOverheadPct": 6, "wallClockMinutesPerTask": 18, "contextCompaction": "disabled" }, "usage": { "runsPerDay": 22000, "delta7dPct": 24, "basis": "estimate" }, "stack": { "frameworks": [ "synthesia-video-agent" ], "models": [ "GPT-6.1 Sol", "ElevenLabs voice stack" ], "tools": [ "avatar-engine", "tts", "translation", "captioning", "brand-kit" ], "runtime": "Synthesia cloud render farm" }, "workflow": [ { "step": "Parse source", "detail": "Extracts structure from the script, deck or document", "tokensPct": 12, "toolCalls": 1 }, { "step": "Plan scenes", "detail": "Splits into scenes with b-roll and on-screen text", "tokensPct": 16, "toolCalls": 1 }, { "step": "Write script", "detail": "Rewrites for spoken delivery and pacing", "tokensPct": 28, "toolCalls": 2 }, { "step": "Render video", "detail": "Generates avatar performance, voice and captions", "tokensPct": 32, "toolCalls": 3 }, { "step": "Localise", "detail": "Adds language variants and brand checks", "tokensPct": 12, "toolCalls": 1 } ], "deliverables": [ "Rendered MP4 video", "Scene-by-scene script", "Localised language variants" ], "links": { "website": "https://www.synthesia.io", "github": "", "paper": "" }, "tags": [ "marketing", "video", "avatar", "localisation" ], "metricsNote": "Estimated: seat-plus-minute pricing; per-video token spend modelled from a 6-minute training video with three localisations." }, { "slug": "julius", "name": "Julius AI", "tagline": "Data analyst agent that runs real Python on your spreadsheet", "description": "Julius takes a dataset or spreadsheet, inspects the schema, writes and executes Python for the analysis, and returns charts plus a written interpretation. Because the code actually runs in a sandbox, the agent can iterate on failed cells and export a reproducible notebook instead of guessing at numbers.", "domain": "data", "icon": "๐Ÿ“Š", "featured": true, "sponsored": false, "status": "production", "verified": false, "source": { "type": "docs", "label": "Julius AI โ€” AI data analyst pricing and product docs", "url": "https://julius.ai/pricing" }, "metrics": { "tokensPerTask": 85000, "costPerTaskUsd": 0.55, "successRatePct": 86, "toolCallsPerTask": 16, "retryOverheadPct": 6, "wallClockMinutesPerTask": 6, "contextCompaction": "enabled" }, "usage": { "runsPerDay": 68000, "delta7dPct": 16, "basis": "community-report" }, "stack": { "frameworks": [ "julius-agent" ], "models": [ "Claude Sonnet 5.5", "GPT-6.1 Sol" ], "tools": [ "python-sandbox", "pandas", "matplotlib", "spreadsheet-io", "notebook-export" ], "runtime": "Julius cloud sandbox" }, "workflow": [ { "step": "Inspect data", "detail": "Reads schema, types and row counts", "tokensPct": 14, "toolCalls": 2 }, { "step": "Plan analysis", "detail": "Chooses the statistical approach and charts", "tokensPct": 16, "toolCalls": 2 }, { "step": "Write code", "detail": "Generates pandas and plotting code", "tokensPct": 30, "toolCalls": 4 }, { "step": "Execute and fix", "detail": "Runs the code, reads errors, patches cells", "tokensPct": 26, "toolCalls": 6 }, { "step": "Interpret", "detail": "Writes findings and exports the notebook", "tokensPct": 14, "toolCalls": 2 } ], "deliverables": [ "Analysis notebook", "Charts and figures", "Written interpretation", "Cleaned dataset export" ], "links": { "website": "https://julius.ai", "github": "", "paper": "" }, "tags": [ "data", "analytics", "python", "spreadsheet" ], "metricsNote": "Estimated: Julius sells plans with message limits; per-analysis tokens modelled from a 10k-row CSV session." }, { "slug": "hex-magic", "name": "Hex Magic", "tagline": "Notebook agent that writes, runs and explains SQL and Python", "description": "Hex Magic works inside a shared notebook: it writes SQL against the warehouse, switches to Python for modelling, runs both, and explains each cell in place. Analysts keep the notebook as the deliverable, so the agent's output is reviewable code rather than a chat answer.", "domain": "data", "icon": "๐Ÿ”ฎ", "featured": false, "sponsored": false, "status": "production", "verified": false, "source": { "type": "docs", "label": "Hex โ€” Magic AI for data notebooks", "url": "https://hex.tech/product/magic/" }, "metrics": { "tokensPerTask": 120000, "costPerTaskUsd": 0.75, "successRatePct": 83, "toolCallsPerTask": 20, "retryOverheadPct": 8, "wallClockMinutesPerTask": 9, "contextCompaction": "enabled" }, "usage": { "runsPerDay": 24000, "delta7dPct": 10, "basis": "estimate" }, "stack": { "frameworks": [ "hex-magic" ], "models": [ "Claude Sonnet 5.5", "GPT-6.1 Sol" ], "tools": [ "warehouse-sql", "python", "dbt", "chart-cells", "semantic-layer" ], "runtime": "Hex cloud notebook" }, "workflow": [ { "step": "Read schema", "detail": "Inspects warehouse tables, joins and lineage", "tokensPct": 16, "toolCalls": 3 }, { "step": "Write SQL", "detail": "Drafts the query and checks row counts", "tokensPct": 26, "toolCalls": 5 }, { "step": "Validate", "detail": "Explains the result and flags anomalies", "tokensPct": 20, "toolCalls": 4 }, { "step": "Model in Python", "detail": "Switches to Python for stats or ML", "tokensPct": 24, "toolCalls": 5 }, { "step": "Document", "detail": "Adds cell explanations and a summary", "tokensPct": 14, "toolCalls": 3 } ], "deliverables": [ "Notebook with SQL and Python cells", "Charts", "Cell-level explanations" ], "links": { "website": "https://hex.tech", "github": "", "paper": "" }, "tags": [ "data", "sql", "notebook", "warehouse" ], "metricsNote": "Estimated: Hex bills per seat plus AI credits; per-session tokens modelled from a 6-cell analysis notebook." }, { "slug": "deep-research", "name": "OpenAI Deep Research", "tagline": "Long-horizon research agent that browses and writes a cited report", "description": "Deep Research plans a multi-step investigation, browses dozens of pages, cross-checks sources and returns a structured report with citations in 5-30 minutes. It is priced as a premium ChatGPT tier or per-query API spend, which puts a single deep query at several dollars of inference.", "domain": "data", "icon": "๐ŸŒ", "featured": false, "sponsored": false, "status": "production", "verified": false, "source": { "type": "docs", "label": "OpenAI โ€” Introducing deep research (product documentation)", "url": "https://openai.com/index/introducing-deep-research/" }, "metrics": { "tokensPerTask": 2600000, "costPerTaskUsd": 3.2, "successRatePct": 78.2, "toolCallsPerTask": 45, "retryOverheadPct": 12, "wallClockMinutesPerTask": 25, "contextCompaction": "enabled" }, "usage": { "runsPerDay": 40000, "delta7dPct": 33, "basis": "community-report" }, "stack": { "frameworks": [ "deep-research-harness" ], "models": [ "GPT-6.1 Sol", "GPT-6 Astra" ], "tools": [ "web-browser", "python", "file-reader", "vector-store" ], "runtime": "OpenAI managed cloud" }, "workflow": [ { "step": "Clarify scope", "detail": "Restates the question and asks follow-ups", "tokensPct": 6, "toolCalls": 2 }, { "step": "Plan searches", "detail": "Plans the search tree and source types", "tokensPct": 10, "toolCalls": 4 }, { "step": "Browse", "detail": "Reads pages, PDFs and datasets, extracts evidence", "tokensPct": 36, "toolCalls": 22 }, { "step": "Cross-check", "detail": "Verifies claims against multiple sources", "tokensPct": 24, "toolCalls": 10 }, { "step": "Write report", "detail": "Produces the cited long-form report", "tokensPct": 24, "toolCalls": 7 } ], "deliverables": [ "Cited research report", "Source table", "Follow-up question list" ], "links": { "website": "https://openai.com/index/introducing-deep-research/", "github": "", "paper": "" }, "tags": [ "deep-research", "web-navigation", "openai" ], "metricsNote": "Estimated: OpenAI does not publish per-query tokens; usage modelled from a 40-source report at o-series pricing." }, { "slug": "n8n", "name": "n8n AI agents", "tagline": "Workflow automation platform where LLM agents call real integrations", "description": "n8n lets teams build agent workflows as node graphs: an LLM node reasons over incoming events, calls HTTP and database nodes, and branches on the result, with the full execution trace kept per run. It is self-hostable, so the marginal cost of a run is the model tokens plus the workflow's own overhead.", "domain": "ops", "icon": "๐Ÿ”—", "featured": true, "sponsored": false, "status": "open-source", "verified": true, "source": { "type": "community", "label": "n8n issue #38130 โ€” AI Assistant run: ~137k input tokens and $0.42 for a five-word question", "url": "https://github.com/n8n-io/n8n/issues/38130" }, "metrics": { "tokensPerTask": 137000, "costPerTaskUsd": 0.42, "successRatePct": 74, "toolCallsPerTask": 12, "retryOverheadPct": 15, "wallClockMinutesPerTask": 3, "contextCompaction": "disabled" }, "usage": { "runsPerDay": 150000, "delta7dPct": 20, "basis": "community-report" }, "stack": { "frameworks": [ "n8n" ], "models": [ "GPT-6 Luna", "DeepSeek V4.1 Flash", "GLM-4.7" ], "tools": [ "http-request", "postgres", "slack", "webhook", "mcp-servers" ], "runtime": "Self-hosted Node.js or n8n cloud" }, "workflow": [ { "step": "Trigger", "detail": "Webhook, schedule or app event starts the workflow", "tokensPct": 6, "toolCalls": 1 }, { "step": "Agent turn", "detail": "LLM decides which tool node to call", "tokensPct": 34, "toolCalls": 4 }, { "step": "Tool nodes", "detail": "HTTP, SQL and app nodes execute the action", "tokensPct": 30, "toolCalls": 4 }, { "step": "Branch", "detail": "Conditional logic and error handling paths", "tokensPct": 18, "toolCalls": 2 }, { "step": "Write back", "detail": "Updates the CRM, sheet or database and logs the run", "tokensPct": 12, "toolCalls": 1 } ], "deliverables": [ "Automated workflow", "Execution log with per-node tokens", "Integration side effects" ], "links": { "website": "https://n8n.io", "github": "https://github.com/n8n-io/n8n", "paper": "" }, "tags": [ "ops", "automation", "workflow", "self-hosted" ], "metricsNote": "Per-run tokens and $0.42 cost from a reported n8n AI Assistant run; success rate estimated from community workflow reports." }, { "slug": "zapier-agents", "name": "Zapier Agents", "tagline": "No-code agents that take actions across 7,000+ integrations", "description": "Zapier Agents wrap an LLM around Zapier's integration graph: the agent interprets a request, picks the right app actions and executes them, with human approval steps where configured. Because every action is a Zapier task, cost per run is a mix of AI credits and task quota.", "domain": "ops", "icon": "โšก", "featured": false, "sponsored": false, "status": "production", "verified": false, "source": { "type": "docs", "label": "Zapier โ€” Agents product documentation", "url": "https://zapier.com/agents" }, "metrics": { "tokensPerTask": 45000, "costPerTaskUsd": 0.35, "successRatePct": 76, "toolCallsPerTask": 8, "retryOverheadPct": 9, "wallClockMinutesPerTask": 4, "contextCompaction": "disabled" }, "usage": { "runsPerDay": 88000, "delta7dPct": 7, "basis": "estimate" }, "stack": { "frameworks": [ "zapier-agents" ], "models": [ "GPT-6 Luna", "Claude Sonnet 5.5" ], "tools": [ "zapier-integrations", "webhooks", "tables", "approval-steps" ], "runtime": "Zapier cloud" }, "workflow": [ { "step": "Parse request", "detail": "Reads the trigger event and the agent instructions", "tokensPct": 12, "toolCalls": 1 }, { "step": "Choose actions", "detail": "Selects the apps and operations to run", "tokensPct": 22, "toolCalls": 2 }, { "step": "Execute", "detail": "Runs Zapier actions across connected apps", "tokensPct": 34, "toolCalls": 3 }, { "step": "Handle errors", "detail": "Retries or routes failures to a human", "tokensPct": 18, "toolCalls": 1 }, { "step": "Confirm", "detail": "Writes the result back and notifies the owner", "tokensPct": 14, "toolCalls": 1 } ], "deliverables": [ "Completed multi-app workflow", "Run history", "Approval records" ], "links": { "website": "https://zapier.com/agents", "github": "", "paper": "" }, "tags": [ "ops", "no-code", "integrations", "automation" ], "metricsNote": "Estimated: Zapier bills task quota plus AI credits; per-run tokens modelled from a three-app automation." }, { "slug": "lindy", "name": "Lindy", "tagline": "Personal ops agent that runs inbox, calendar and CRM chores", "description": "Lindy builds small personal agents that watch an inbox, calendar or CRM, decide what needs doing and take the action: scheduling, follow-ups, data entry and handoffs. It is aimed at operations work where the value is a completed chore rather than a generated document.", "domain": "ops", "icon": "๐Ÿ—‚๏ธ", "featured": false, "sponsored": true, "status": "production", "verified": false, "source": { "type": "docs", "label": "Lindy โ€” AI agents for business operations (plans and pricing)", "url": "https://www.lindy.ai/pricing" }, "metrics": { "tokensPerTask": 60000, "costPerTaskUsd": 0.4, "successRatePct": 72, "toolCallsPerTask": 10, "retryOverheadPct": 12, "wallClockMinutesPerTask": 6, "contextCompaction": "disabled" }, "usage": { "runsPerDay": 12000, "delta7dPct": 29, "basis": "estimate" }, "stack": { "frameworks": [ "lindy-agent-framework" ], "models": [ "GPT-6.1 Sol", "Claude Sonnet 5.5" ], "tools": [ "gmail", "google-calendar", "hubspot", "slack", "webhooks" ], "runtime": "Lindy cloud" }, "workflow": [ { "step": "Watch trigger", "detail": "Email, calendar event or CRM change starts the agent", "tokensPct": 10, "toolCalls": 1 }, { "step": "Decide", "detail": "Reasons about what action the trigger needs", "tokensPct": 24, "toolCalls": 2 }, { "step": "Act", "detail": "Drafts the reply, books the slot or updates the CRM", "tokensPct": 34, "toolCalls": 3 }, { "step": "Check", "detail": "Confirms the outcome against the instruction", "tokensPct": 18, "toolCalls": 2 }, { "step": "Report", "detail": "Notifies the owner and logs the run", "tokensPct": 14, "toolCalls": 2 } ], "deliverables": [ "Completed chore", "Drafted replies and bookings", "Run log with approvals" ], "links": { "website": "https://www.lindy.ai", "github": "", "paper": "" }, "tags": [ "ops", "personal-agent", "inbox", "scheduling" ], "metricsNote": "Estimated: seat-plus-credit pricing; per-run usage modelled from a 3-step inbox chore." } ] }