The loop with a plan and a checklist

Manus runs the now-familiar loop, framed as analyze, plan, execute, observe, repeated until the task is done, with one deliberate constraint: one tool action per iteration. It must wait for each result before deciding the next step. This prevents the model from firing off a long chain of unchecked operations and keeps every step observable, which matters a lot for an agent that runs unattended.

Two modules structure the autonomy:

• A planner module decomposes a high-level goal into an ordered list of steps, injected as a "Plan" event the agent follows and can revise on the fly. This is the orchestrator's planner from Chapter 11, built into a single agent.
• A knowledge module and a datasource module inject reference material and authoritative data (via pre-approved APIs) into context when relevant, so the agent prefers real data over web scraping. This is retrieval-augmented generation wired into the loop, and it is the anti-hallucination contract from Chapter 10 ("prefer authoritative sources; do not make things up").