# SWE-agent

SWE-agent publishes no per-task cost or token figure in the AgentBurn index. AgentBurn snapshot 2026-10-27. Every figure carries a provenance label: source-backed, derived, or left empty when no public basis exists.

| Metric | Value | Basis |
|---|---|---|
| Tokens per task | not published | not_public |
| Cost per task | not published | not_public |
| Success rate | not published | not_public |
| Cost per solved task | not published | — |
| Runs per day (estimate) | not published | not_public |

Princeton agent with a bespoke computer interface for repository tasks

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.

Data snapshot: 2026-10-27 · https://getagentburn.com/projects/swe-agent/