Vercel deleted 80% of its agent’s tools and the agent got better + what to delete from yours (guide inside!)

Source: Nate’s Newsletter Date: 2026-06-17 URL: https://natesnewsletter.substack.com/p/ai-agent-maintenance

Summary

A Nate’s Newsletter piece (paywalled in full) reports that Vercel stripped roughly 80% of the tools from one of its production agents and observed improved performance. The visible framing is that tool accumulation is the primary failure mode in deployed agents — “more context, more tools, more memory feels like care; usually it is the thing rotting your agent from the inside.” The piece promises a seven-part audit taxonomy for which agent components go stale, with tools listed as the first failure class.

Implications

• Native-tooling speed. The result runs counter to the prevailing instinct to give agents maximum capability surface. Fewer, sharper tools beats comprehensive tool registries — which has direct bearing on how agent platforms (Vercel AI SDK, LangChain, OpenAI Assistants) should think about tool registration APIs. More permissive registration makes stale-tool rot more likely, not less.
• Agent-form / proactivity. Tool bloat degrades not just performance but agent confidence: an agent uncertain which of fifteen tools to call stalls or calls the wrong one. The Vercel finding is a concrete data point for the narrower-scope, higher-reliability end of the agent design spectrum — closer to purpose-built than generalist.
• Two-clocks (open capability). The open-source tooling community has been moving toward maximally composable, plug-any-tool registries (MCP, tool-use APIs). Vercel’s finding pushes back: unbounded composability is a footgun in production. Watch whether this surfaces as guidance in MCP server spec revisions or agent framework defaults.