Stop Giving Your Agent Tools. Give It a Sandbox - Mikyo King, Arize AI
Give agents a code sandbox instead of tool catalogs to eliminate context bloat
“Essentially, whatever you need, the coding agent can actually create the software for you dynamically.”
Arize AI's Michael King argues that the dominant agent pattern—giving an LLM a large static catalog of tools—causes context bloat, irrelevant tool invocations, and compaction problems at scale. His proposed alternative, 'Code Mode,' lets the agent dynamically write and execute the code it needs rather than selecting from a pre-defined menu, aligning agent design with how frontier models are actually trained (e.g., SWE-bench). This reframes agent tooling as a sandbox architecture problem, with implications for how practitioners design long-horizon, multi-turn agents.