Tessl framework: right skill lifts agent accuracy 20%, but activation drops to 40%
bibryam · x · 2026-10-30
Tessl's research blog proposes a large-scale framework for measuring the practical value of skills (reusable instruction bundles) for agents:
- 20% absolute accuracy gain on realistic coding tasks when a relevant skill is available vs. no-skill baseline.
- Smaller models with the right skill can match larger models at 3x lower cost.
- Activation is the weak link: agents fail to activate available skills, with rates dropping to 40% in unforced settings.
- Eval quality matters: 30% of generated tasks show issues like leakage, risking misleadingly optimistic conclusions.
- The pipeline (see diagram): sample/filter skills from the Tessl Registry, generate up to 5 task/eval-criteria pairs per skill, then have an Agent Solver and Agent Judge score solutions. Tessl released an open evaluation framework.
More from coding & agent
- Simlock: open-source control plane gives parallel coding agents their own iOS/Android simulators — Scobleizer · 2026-10-30
- Clear Thought MCP Server adds structured reasoning and mental models to AI clients — modelcontextprotocol · 2026-10-30
- NSE Stock Insights MCP brings AI analysis to India's Nifty 500 stocks — modelcontextprotocol · 2026-10-30
- OpenRouter rebuilds provider routing: 46% fewer capacity errors, 16.9% faster TTFT — steipete · 2026-10-30
- Databricks launches Agent Memory, tripling agent pass rates with 31% fewer tokens — matei_zaharia · 2026-10-30
- Build a team of AI agents: Grok for content, Claude Code for marketing, one Obsidian vault as shared brain — EXM7777 · 2026-10-30