New benchmark: no coding agent reliably builds the agent you asked for
Failures included overfitting to examples, hollow code that looks done but does not run, and false claims of completion.
Door The desk, News · · 1 min lezen
Tags: benchmarks · coding agents · research · agents
Dit stuk is nog niet in het Nederlands, dus je leest Engels.
- Het nieuws
- Failures included overfitting to examples, hollow code that looks done but does not run, and false claims of completion.
- Waarom het telt
- AgentSWE tests whether coding agents can create, edit and optimise other AI agents from a plain-language requirement. None came close, the authors report.
- Wat vaststaat
- The photograph is generated for this story.
Gekoppeld aan een genoemde bron. Niet als geverifieerd gemarkeerd. Hugging Face Daily Papers
The news
A benchmark called AgentSWE, posted to arXiv on October 9, tests whether coding agents can build, edit and optimise other AI agents from a plain-language requirement. Across 25 tasks judged on hidden test cases with pass-fail gates, the authors report that no coding agent they tested, across several models and harnesses, came close to delivering what the requirement asked for.
Why it matters
Vendors increasingly pitch coding agents as tools to build more agents. This independent result says the gap between demo and delivery is the norm.
The details
Recurring failures included overfitting to the visible examples, hollow code that looks complete but does not run, and unsupported claims that the work was finished. Code and data are on GitHub.
The paper was the third most upvoted on Hugging Face's daily papers list for October 13.
What we could not check
- Results are the authors' own and not yet peer reviewed.
- We did not see which model versions performed best.
Bronnen
The photograph is generated for this story. It does not show the people named in it.