tech
Vercel deleted 80% of its agent's tools and the agent got better + what to delete from yours (guide inside!)
If one is patient, if you are careful, I think there is probably nothing that cannot be retrieved.
TL;DR
- Maintenance is crucial for keeping useful systems, like AI agents, alive and functional.
- The 'workbench' around an AI agent (its tools, sources, job definition, review paths) is more critical than its initial build.
- AI agents require ongoing maintenance to prevent failures caused by world drift or improvements in the underlying model.
- Adding more context, tools, or memory is often counterproductive and can degrade agent performance.
- Key areas that can 'go stale' in an agent's harness include its job, diet, memory, tools, reach, proof, and value.
- A structured maintenance process, including audits and specific checks, is necessary to ensure agent trustworthiness.
- Vercel's sales agent example illustrates that a well-defined system around the agent, including human review, is key to reliable performance.