Key takeaways

• Manus is a general autonomous agent in a cloud sandbox: a foundation model plus a planner, tools, retrieval, and file memory, producing real deliverables end to end.
• CodeAct makes the model's action a short Python program rather than a fixed tool call, which is far more expressive and self-correcting, at the cost of needing tight sandboxing.
• The loop is analyze, plan, execute, observe, with one action per iteration so every step is observable and controllable.
• File-based memory beats the context limit: write intermediate results to files, keep a live todo.md checklist, and draft long documents in pieces. The checklist is the source of truth if context is lost.
• The architecture is replicable with open tools; the hard part is reliability (error handling, loop prevention, not fabricating), not the architecture itself.

Original sources: the Manus technical analysis and replication blueprint (based on the leaked Manus prompt and the CodeAct ICML 2024 paper "Executable Code Actions Elicit Better LLM Agents").

---