CH 14 · Autonomous Task Agents: Manus and CodeAct · 08/11
Read as one pageHow it works: a CodeAct loop with a planner and a todo file
The MVP combines three pieces: a planner that turns a goal into a todo.md, a CodeAct executor that runs the model's Python in a (here, simplified) sandbox and feeds back the result, and file-based memory that persists progress. One action per iteration, observe, repeat.
Finished this idea?
Your learning progress stays in this browser.