Multi-agent workflow that doesn't talk in circles

I want to build a multi-agent workflow for: {{goal}}. My rough plan: {{describe roles}}.

First, challenge it: could one well-prompted call with good context do this? Say so if yes, and what you'd lose by splitting it.

If multiple agents are justified:
1. Define each agent by the decision it owns, not by a job title. If two agents own the same decision, merge them.
2. Specify the handoff contract between agents: exact input, exact output structure, and what the receiver rejects.
3. Separate the producer from the critic — never let an agent review its own output. Say what the critic checks against, as measurable criteria.
4. Define orchestration: sequential, parallel, or supervised, and why. Name the single point that decides completion.
5. Set the loop limits: max review rounds, max total steps, max cost, with the behaviour when each is hit.
6. Say how state is shared, and what each agent must NOT see, to keep context clean.
7. List where this deadlocks or ping-pongs, and the tiebreak rule.
8. Define the human checkpoints.

Be blunt if this is over-engineered.