Vercel cuts agent tools 17 to 2; ETH Zurich finds AGENTS.md adds cost

The discussion started from a claim that overloading the harness makes models underperform, which prompted the author to reconsider the many rules and skills added to their own setup. The reports cited point in a consistent direction: Vercel's reduction of an analytics agent's tools from 17 to 2 coincided with more successes, less time and fewer tokens on its 5-question test, and ETH Zurich's study found that handing over context files such as AGENTS.md did not generally lift task success while raising inference cost by over 20% on average. A Japanese talk by Gota at Harness Engineering Meetup Tokyo #1 in April 2026 described settings files exceeding 300 lines, where changing one task template required edits in 10 places and a full test run took an hour, making it hard to tell which fix mattered.

The author's explanation is that an AI reads everything it is given and tries to obey all of it, so unrelated rules dilute the important ones, similar tools cause hesitation, and small contradictions between rules cause stalls or detours. A concrete example is a rule requiring tests after code changes colliding with a later rule forbidding DB connections, which can lead the agent to skip tests or rewrite them. The proposed split is between instructions, which the AI may or may not follow, and enforcement mechanisms like hooks, permissions, tests and lint, which act regardless of the AI's judgment; overloading problems come mostly from instructions, while enforcement does not confuse the model. In the author's repository, CLAUDE.md went from 102 lines to 50, with rules moved to .claude/rules/ and media-specific information placed in folder-level CLAUDE.md files and skills, and the author says they will write later about how that worked out.