Twenty-five skills in, your agent is averaging out their conflicts and handing you duller work than it did at five.

Source: Nate’s Newsletter Date: 2026-08-01 URL: https://natesnewsletter.substack.com/p/agent-skill-one-job-test

Summary

The companion claim to the “one-job test” piece: past a threshold, stacking skills degrades output because the model has a bounded token budget for skill visibility and because conflicting process opinions get averaged into blander, lowest-common-denominator work. Twenty-five skills can underperform five.

Implications

Feeds the context-economics / “more isn’t better” radar thread. This is the skill-layer instance of a pattern showing up across the stack: agent quality is capped by working-set discipline, not additive with installed capability. Connects to the token-cost-as-operating-cost thread (every loaded skill spends context budget that could be reasoning) and to the harness-size guidance already tracked (CC’s <15-agent default, the affected-set / “hold less” instinct visible in tooling). The design lesson: curate and test skills like dependencies, prune aggressively, and treat a growing skill library as accruing conflict debt rather than compounding leverage.