SWE-agent Hold
Overview
SWE-agent is a research system that solves GitHub issues by interacting with repositories and test harnesses (SWE-agent). It introduced the idea of an agent-computer interface tailored to language models and reported 12.5% pass@1 on SWE-bench in its 2024 paper, state of the art at the time (arXiv). The project is MIT-licensed and maintained by researchers from Princeton and Stanford, who now state that the simpler mini-swe-agent has superseded it (GitHub).
SWE-agent is rated Hold because its own maintainers have moved development to mini-swe-agent and recommend it instead. Hold applies to new work on SWE-agent itself; mini-swe-agent is the maintained successor for research and benchmark-harness use.
Adoption Signals
- The SWE-agent README carries a warning that "mini-swe-agent … has superseded SWE-agent" and recommends mini-swe-agent "going forward"; the last tagged release is v1.1.0 (22 May 2025), and commits since then are small fixes and documentation changes (GitHub, commit history).
- mini-swe-agent is actively released (v2.4.6 on 23 July 2026, about 8,100 stars on 30 September 2026) and its maintainers list Meta, NVIDIA, Essential AI, IBM, Nebius, Anyscale, Princeton, and Stanford as users (mini-swe-agent).
- mini-swe-agent is the harness behind the SWE-bench bash-only leaderboard, which compares models under the same minimal agent (mini-swe-agent, SWE-bench).
- OpenAI stopped reporting SWE-bench Verified after its audit and recommends SWE-bench Pro, which weakens the benchmark results SWE-agent is known for (AI Weekly, Decrypt).
Risks
- Building on a maintenance-only line. Internal tooling on SWE-agent inherits a codebase whose authors put their development effort into mini-swe-agent; there has been no tagged release since May 2025, and the main branch only receives small fixes (GitHub, commit history), so support for new model APIs is not planned work.
- Benchmark validity. OpenAI's audit of 138 SWE-bench Verified tasks that its models repeatedly failed found 59.4% flawed, and models reproduced original fixes from memory (AI Weekly, Decrypt); published scores say little about performance on a company's own code.
- Execution of model-generated commands. Both agents run shell commands proposed by the model; mini-swe-agent documents local execution alongside Docker, Podman, and bubblewrap (mini-swe-agent), and running locally gives the model the developer's shell and credentials.
- No governance layer. Neither project provides approvals, per-user permissions, or audit logs, so wiring either into issue trackers for automatic fixes bypasses code review unless that is built around it.
Pros & Cons
Advantages
- SWE-agent is a well-known, MIT-licensed research system from the Princeton and Stanford team behind SWE-bench, configured through a single YAML file, so it looks like a safe, documented starting point for agent experiments.
- Its agent-computer interface paper shaped later coding agents, and SWE-agent 1.0 with Claude 3.7 reported state-of-the-art results on SWE-bench Verified in February 2025, which still gives it name recognition.
- About 20,400 GitHub stars and existing tutorials make it the first result many teams find when they look for an open-source issue-fixing agent.
Disadvantages
- The maintainers state that mini-swe-agent has superseded SWE-agent and recommend using it instead; the last tagged SWE-agent release is v1.1.0 from 22 May 2025.
- It was built for research runs, not team workflows, and has no review gates, access management, or audit trail comparable to maintained commercial or self-hosted coding agents.
- The benchmark it is known for has lost credibility: an OpenAI review of 138 SWE-bench Verified tasks its models kept failing found 59.4% flawed, plus evidence of training data contamination.
Recommendation
Hold SWE-agent for new work: do not start new experiments, integrations, or internal tools on it. Existing SWE-agent setups should be migrated to mini-swe-agent, which the same team maintains and recommends; its FAQ explains the differences between the two (mini-swe-agent FAQ).
For research and evaluation — comparing models or prompts on historical issues from your own repositories, or teaching teams how coding agents work — use mini-swe-agent in containers without production credentials, and rely on held-out internal tasks rather than public SWE-bench Verified scores. For day-to-day coding work, use the maintained agents on this radar instead: OpenHands (Trial) as a self-hosted open-source option, and Claude Code (Adopt) or GitHub Copilot (Adopt) as managed tools with review and access controls.
Sources
Consulting
From radar to implementation
You have assessed the technologies — we help you bring them into production systems safely and measurably.
