CH 14 · Autonomous Task Agents: Manus and CodeAct · 01/11
Read as one pageAutonomous Task Agents: Manus and CodeAct
CodeAct has the model write a short Python program as its action instead of a fixed function call. Manus keeps a todo.md file as memory.
11 slides · 6 min read
In this chapter
- 01The setup: a digital worker in a sandbox
- 02CodeAct: code as the action language
- 03The loop with a plan and a checklist
- 04File-based memory: the todo.md trick
- 05Replicating it with open tools
- 06Why it matters
- 07How it works: a CodeAct loop with a planner and a todo file
- 08Code MVP: a CodeAct-style agent loop
- 09Connecting to the bigger picture
- 10Key takeaways
Coding agents like Claude Code and Codex are tuned for software work in a repo. Manus is a more general autonomous agent: give it a goal ("research this topic and write a report," "build and deploy a small website") and it runs end to end in a cloud sandbox, browsing the web, writing and running code, and producing real deliverables. It is a great capstone case study because it combines almost everything from earlier chapters (the loop, tools, sandbox, memory, planning, multi-agent) and adds two distinctive ideas worth stealing: CodeAct and file-based memory. Most of what follows comes from a reverse-engineering report based on Manus's leaked system prompt and the CodeAct research paper it cites.
Finished this idea?
Your learning progress stays in this browser.