Rating · evidence to 26 Sept 2026
Not enough evidence
- Trust B
- Open source
The Register / Coding agents / SWE-agent / mini-SWE-agent
Register entry · Coding agents · updated 26 Sept 2026
SWE-agent / mini-SWE-agent
SWE-agent team (Princeton University & Stanford University) · Coding agents
SWE-agent is a Princeton/Stanford research agent that lets a language model fix issues in GitHub repositories through a custom agent-computer interface. Its successor, mini-SWE-agent, reduces the agent to roughly 100 lines of Python that give the model only bash, and the team now recommends it by default. mini-SWE-agent is used as the reference harness on SWE-bench bash-only and several Terminal-Bench and third-party leaderboards, and offers a small interactive CLI (`mini`).
What reviewers and users praise
- Very small, readable codebase that is easy to audit and modify
- Strong benchmark results for its size; used as a neutral harness by labs and leaderboards
- Works with essentially any model because it needs only text output
What they criticise
- Minimal feature set compared with product harnesses (no IDE integration, limited UX)
- Not beginner-friendly: users manage Python environments, model config and sandboxing
- Older SWE-agent has an unfixed high-severity CVE in its trajectory inspector
Experts vs the crowdneeds both scores
Experts—
Crowd—
| Price | Free |
|---|---|
| Use it via | CLI, API |
| Runs on | macos, linux, windows |
| Licence | MIT |
| Updated | Evidence collected 26 Sept 2026 |
How the rating was worked out
Weights for Coding agents: adoption 35%, experts 40%, crowd 25%. Fewer than two pillars have enough evidence. The method.
| Adoption & momentumweight 35% | 70 | GitHub stars 8k → 58 · PyPI monthly downloads 10M → 92 · 4 releases in 90 days → 60reach × 0.7 + momentum × 0.3 |
|---|---|---|
| Expertsindependent reviews · weight 40% | — | Only 2 independent reviews (3 needed) Benchmarks, shown not scored: Terminal-Bench 2.0 No. 40 of 43 (Mini-SWE-Agent + Claude Sonnet 4.5) · SWE-bench Verified No. 4 of 77 (mini-SWE-agent + Claude 4.5 Opus) · SWE-bench Lite No. 4 of 58 (SWE-agent + Claude 4 Sonnet) · SWE-bench Test No. 5 of 13 (SWE-agent 1.0 + Claude 3.7 Sonnet) · SWE-bench Multimodal No. 8 of 12 (SWE-agent + Claude Sonnet 3.5) |
| Crowdcommunity sample · weight 25% | — | Community sample from a single platform |
| Rating | — | Not rated: fewer than two pillars have enough evidence. |
Evidence checks. 1 review from tool directories, AI-written pages or unnamed authors not counted. 2 GitHub issues not counted as sentiment. How evidence is checked.
Editorial adjustment. Incident-history grade reviewed under methodology revision 2 (see the trust section).
Trust & safety
Seven dimensions graded A–D from documented facts. Overall: B.
B
Permission model
`mini` CLI has confirm mode (user approves each proposed command), yolo mode (-y, executes immediately) and human mode; the interactive CLI runs in the local environment without a sandbox. Docker/Podman/Singularity environments exist for batch runs. source
B
Data access scope
Agent acts only through bash in the chosen environment, so it has the full access of the shell user (or of the container when one is used); documented. source
A
Data storage
Runs locally; no vendor service. Trajectories are saved locally. source
n/e
Data retention & training
No vendor-hosted service; no telemetry statement found in docs. Retention depends on the user's model provider. source
D
Incident history (24 months)
CVE-2026-75482 (CVSS 3.1 7.5 HIGH) affects SWE-agent v1.1.0's trajectory inspector; the reporting issue is still open and no SWE-agent release has shipped since. No CVEs found for mini-SWE-agent itself. Kept at D under revision 2: the reporting issue for CVE-2026-75482 is still open and no SWE-agent release has shipped since (unresolved). source
D
Compliance
Academic open-source project; no attestations (N/A). source
A
Transparency
MIT-licensed with public GitHub release notes. source
The experts
2 named independent reviews found; dated reviews from the last 12 months are scored.
- positive
Research-focused tool with innovative Agent-Computer Interface and minimal footprint
LocalAIMaster · 6 Feb 2026 · localaimaster.com - positive
the radically simple, research-grade choice
Rigor (DEV Community) · 3 Sept 2026 · dev.to
The crowd
7 coded posts from hackernews.
5 positive2 mixed0 negative
- positive
The harness can be almost barebones (like, for example, mini-swe-agent used for benchmarks), and yet the model still does the task just fine.
Hacker News comment · 20 Jul 2026 · news.ycombinator.com - mixed
They choose it because it has features that make it behave much more nicely as a pair programming assistant than mini-swe-agent.
Hacker News comment · 8 Jan 2026 · news.ycombinator.com - positive
Personally I found the mini SWE-agent to be a very approachable introduction to building agents
Hacker News comment · 4 Oct 2025 · news.ycombinator.com - positive
See also: agent harness in 50 lines (based on mini-swe-agent).
Hacker News comment · 7 Jun 2026 · news.ycombinator.com - mixed
does anyone know why they used "mini-SWE-agent" not Claude Code? I have never heard of this agent before
Hacker News comment · 25 Jul 2026 · news.ycombinator.com - positive
We hypothesize that the superior performance of Opus 4.7 on mini-swe-agent relative to the more complex Claude Code harness stems from the tight feedback loop (edit-run-check)
Hacker News comment · 15 May 2026 · news.ycombinator.com - positive
is way better today than the best agentic harness of last year. Check out mini-swe-agent.
Hacker News comment · 14 Jan 2026 · news.ycombinator.com
Pricing
As published, checked 26 Sept 2026.
| Plan | Price | Notes |
|---|---|---|
| Open source | Free | MIT-licensed; user pays their own model costs. |
Recent changes
- 3 Sept 2026
Updated ProgramBench task prompt on mini-SWE-agent main. source
- 23 Jul 2026
mini-SWE-agent v2.4.6: failed-to-parse billed calls now count against cost_limit. source
- 6 Jul 2026
mini-SWE-agent v2.4.5: config files read as UTF-8 regardless of locale; model error handling fixes. source
- 16 Jul 2026
SWE-agent maintenance fixes (repo URL parsing, multimodal subset mapping); no new release. source
Ranked alongside
Other ranked entries in Coding agents.
OpenCodeAnomaly75
Claude CodeAnthropic73
Zed (Agent Panel)Zed Industries73
GitHub CopilotGitHub (Microsoft)72
Codex (CLI and app)OpenAI71Sources
Every figure above traces to one of these, observed up to 26 Sept 2026.