Crabhaus

What it is

A pattern where the agent's action is a program it writes and runs, not one tool call at a time. Instead of emitting JSON calls and passing every result through its context window, the model writes a script that calls many tools, loops, and filters data in a sandbox - only the final result returns to context.

How it works

Tools are exposed as a code API - each MCP server becomes an importable module. The model loads only the definitions it needs (progressive disclosure), writes code chaining the calls, and a sandboxed interpreter runs it. Intermediate data stays in the execution environment and is filtered before anything re-enters context - a 10,000-row sheet becomes the five rows that matter.

A concrete example

Anthropic's worked example is a two-system chore: pull a meeting transcript from Google Drive, then file it on a Salesforce lead. As tool calls, the transcript flows through the model twice. As code, the Drive read feeds the Salesforce write inside the sandbox - a 150,000-token interaction shrinks to about 2,000, a 98.7% cut.

Why it matters

The CodeAct paper measured up to 20% higher task success across 17 LLMs when agents act via executable Python instead of JSON calls; Anthropic and Cloudflare ('Code Mode') converged on the pattern to cut cost and latency. The bill is security: running model-written code demands sandboxing, resource limits, and egress controls - this week's sandbox-escape headlines are that failure class.