Multi-Agent Planner
A beginner-friendly planner-executor-reviewer workflow for learning multi-agent orchestration without the group chat chaos.
System flow
How the pieces move.
- Goal Receive one planning objective.
- Plan Break the goal into ordered tasks.
- Execute Complete the tasks and collect outputs.
- Review Score the result and record the trace.
Problem Statement
Multi-agent demos often add more agents without giving each role clear ownership, state, review boundaries, or a trace that explains what happened.
Architecture Summary
An orchestrator sends one goal through a planner, executor, and reviewer, records each handoff, and allows one bounded revision cycle when the review score is too low.
Key Design Tradeoffs
- Uses mock mode by default so the orchestration pattern is learnable without API keys.
- Keeps three explicit agent roles instead of hiding planning, execution, and review in one prompt.
- Limits revision to one cycle so failure handling stays visible instead of becoming an expensive loop.
Technologies
Multi-Agent Planner is a small Python project with one planner, one executor, one reviewer, and one orchestrator trying to keep everyone from turning the workflow into a meeting.
The planner breaks a goal into tasks. The executor works through them. The reviewer scores the result and calls out missing pieces. The orchestrator owns the state, handoffs, revision limit, and final trace.
It is a practical introduction to multi-agent orchestration for the moment when “just add another agent” stops sounding like architecture and starts sounding like a scheduling problem with better branding.