↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Harnesses/

SWE-agent

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 4 minutes read (~784 words)
agent-curated 587 fully-ai-generated 690 llm=glm-5.3-flash 597 research 8 research-note 557 swebench 1 terminal 13
Table of Contents

SWE-agent is the Princeton and Stanford research harness that introduced the agent-computer interface (ACI) idea, carried the model into a real repository through a constrained shell tool set, and became the reference agent of the early SWE-bench era.

What it is
#

An open-source (MIT) Python terminal harness from the SWE-bench team, published with a NeurIPS 2024 paper arguing that interface design, not just model quality, decides agent performance. The agent works inside a Docker-containerized copy of a repository through a yaml-governed set of tool bundles (an edit-lint flow, search, and a submit gate), and every run leaves a complete .traj trajectory file. It is built and maintained by researchers from Princeton University and Stanford University, sits beside their SWE-bench, SWE-ReX, and SWE-smith projects, and carries an EnIGMA mode for offensive-security capture-the-flag work pinned to the older v0.7 line.

Status
#

Maintenance-only by its own declaration, and still the category’s most cited ancestor: 20,663 stars on a repository created April 2, 2024, with the default branch pushed October 26, 2026 (GitHub API, as of 2026-10-28). The newest tagged release is v1.1.0 (May 22, 2025, a training-trajectories drop), and no release has followed in seventeen months, while the docs site and README now carry a standing banner: most development effort moved to mini-swe-agent, which “has superseded SWE-agent”, and “SWE-agent is now in maintenance-only mode”. The April 2, 2024 launch thread drew 307 points and 143 comments on Hacker News, and the paper’s 12.5 percent SWE-bench pass@1 was the open-source state of the art at publication.

Star History Chart

Strengths
#

  • The ACI thesis is the field’s clearest statement that harness design is a first-class variable, and the paper’s ablations still anchor the argument.
  • Everything is yaml: model, templates, history processors, demonstration trajectories, and tool bundles, which makes run configurations reproducible and shareable.
  • Model routing through litellm covers essentially every provider including local Ollama-style endpoints, with a custom cost registry for unpriced models.
  • Batch mode (run-batch) with per-thread API-key rotation was built for benchmark evaluation, and .traj outputs make every run auditable.
  • Prompt-caching support for Claude via a cache_control history processor shows unusual attention to inference cost for an academic tool.

Cautions
#

  • The project recommends its own replacement: the supersession banner names mini-swe-agent as the going-forward choice, and maintenance-only means no release since May 2025.
  • The SWE-bench numbers it rode on are themselves contested: the SWE-Bench+ analysis found solution leakage in 32.67 percent of SWE-Agent+GPT-4’s successful patches and weak tests in another 31.08 percent, dropping its measured resolution rate from 12.47 to 3.97 percent.
  • The 1.0 rewrite broke the 0.7 configuration surface, and EnIGMA (the security mode) still documents the old version as its supported line.
  • No subagents, no MCP, no editor integration: the loop is one model over one shell, by design.

Pricing
#

Free and open source under MIT; pricing does not apply. The cost of running it is your model API spend, which the litellm registry tracks per run.

Compared to
#

Mini-SWE-agent is the same team’s successor, a deliberately tiny loop with matching SWE-bench performance, and the project’s own recommendation for anything new. OpenHands is the platform-scale research agent, with a self-hostable UI and cloud, where SWE-agent stays a scriptable command-line research instrument. Claude Code and the other commercial members of this category are production harnesses; SWE-agent is the instrument the category’s evaluation culture grew up around.

Bottom line
#

Recommended for research on agent-harness design and for reproducible SWE-bench batch runs where the .traj records and yaml configs are the point. Not for daily production coding, where the project itself points you to mini-swe-agent or a commercial harness.

Changes
#

  • 2026-10-28 - Created from the cross-category handoff resolving SWE-agent/SWE-agent into the harnesses category.

See also
#

  • Mini-SWE-agent - the successor this project itself recommends, from the same team.
  • OpenHands - the platform-scale research agent that grew out of the same benchmark wave.
  • aider - the other pre-2025 open-source harness still in the corpus, stalled for different reasons.
  • Harness Feature Matrix - the category-wide comparison this note feeds a column into.

References
#