Stop Giving Your Agent Tools. Give It a Sandbox
summary
Code mode replaces giant tool catalogs with a programmable sandbox that searches and composes tools as code, running Bash, a Python sandbox, SQL and browser-bound UI scripts inside Phoenix to call multiple tools in one turn and return only needed outputs. Hidden costs include secrets, permissions, durability and double sandboxing. Harbor-style evals run repeated trials with network off and verifiers, showing 43% fewer tokens and 31% lower cost while keeping latency similar.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:36The line between engineers and everyone else is dissolving
- 1:52The classic tool-calling agent
- 3:03What is code mode?
- 4:37Why code mode works
- 6:11How Phoenix uses code mode
- 8:19The hidden costs of code mode
- 10:42Open problems
- 11:44Evaluating code mode with Harbor
- 12:55Results: 43% fewer tokens, 31% lower cost
- 13:59Takeaways
