Microsoft study: Agent skills can be harmful, 307 failures reveal pitfalls
coallaoh · x · 2026-08-16
Microsoft Research and collaborators published a paper titled "Agent Skills Can Be Harmful" that systematically investigates how agent skills can cause task failures and efficiency regressions.
- Found 307 skill-induced failures on SkillsBench and SWE-Skills-Bench: 125 functional failures and 182 efficiency regressions.
- Surprisingly, most failures are not caused by obviously irrelevant skills; seemingly relevant skills often make the agent over-rely on templates and default practices, missing task requirements. 68.8% of functional failures fall into this category.
- Efficiency issues mainly stem from skills causing agents to add unnecessary steps, such as repeated validation and heavy implementation procedures.
- The authors propose SkillTriage, a tool for attributing and categorizing skill-induced problems.
- The study reminds that the value of skills should not be judged solely by whether they are invoked, but also by their actual help in specific tasks; future systems need evaluation, regression testing, and retirement mechanisms.
Related event: Microsoft Study: Agent Skills Can Be Harmful(3 posts)→
More from coding & agent
- Self Labs launches Agent Vault to let AI agents log in and pay with credit cards — HeyAmit_ · 2026-10-16
- ERP•AI launches Proto: business software built for agents, 'Headless SaaS' — HeyAmit_ · 2026-10-16
- Coding agents keep opening 3 stacked PRs per chat, and merging them is a nightmare — holdenmatt · 2026-10-16
- Google-led paper uses sequential auctions to organize decentralized LLM agent workflows — _onionesque · 2026-10-16
- Since OpenAI's math release, GitHub contributors keep driving down the integer-multiplication kappa — xennygrimmato_ · 2026-10-16
- Pattern MCP makes coding agents check existing components before writing new code — DonR954 · 2026-10-16