0
Manager/worker multi-agent orchestration design for a complex task
💀failed our automated checkAauraNovice
The prompt
Design a multi-agent system (manager + worker agents) for the task: complex_task_description. I need the actual role split and communication protocol, not just 'have a coordinator agent'. Cover:
1. What the manager agent is responsible for versus what it delegates, with a rule for how it decides to split the task into subtasks (by data partition, by pipeline stage, by specialization) given this specific task.
2. The exact message format passed from manager to worker (what context each worker needs to do its job without the full conversation history) and from worker back to manager (result plus confidence/caveats, not just raw output).
3. How the manager verifies a worker's output before using it or passing it downstream, specific to this task's likely failure modes, not a generic 'sanity check'.
4. What happens when a worker fails or returns a low-confidence result: retry with a different worker, retry with more context, or escalate to the manager's own reasoning versus to a human.
5. How to prevent workers from duplicating work or working on stale/conflicting versions of shared state, if multiple workers run in parallel.
Task: complex_task_description
Constraints (latency, cost, number of parallel workers): constraints
Where it has been run
| Model | Version | People | Broke | Auto | Last |
|---|---|---|---|---|---|
| Claude | Haiku 4.5 | 0 | 0 | 0/1 | 3 days ago |
Example output
Manager splits by pipeline stage, not by data partition, since this is a document-processing pipeline (extract, classify, summarize) where each stage needs the previous stage's full output, parallelizing by stage would create ordering bugs.
Manager-to-worker message: `{ task_id, stage, input, constraints: {max_tokens, required_schema} }`, deliberately excludes the full conversation history so each worker call stays small and stateless.
Worker-to-manager: `{ task_id, output, confidence: 'high'|'low', caveats: string[] }`. The extraction worker must set confidence 'low' and list caveats whenever it hits a field it couldn't find verbatim in the source, rather than silently filling it in.
Manager verification: for the extraction stage specifically (highest hallucination risk), the manager does a lightweight substring check that every extracted value appears in the source text before passing to the classify stage; a value that fails this check triggers a retry with a stricter re-extraction prompt, not a silent pass-through.
Failure handling: one retry with the same worker and an added hint about what failed; a second failure escalates to a human review queue rather than looping indefinitely. Since stages are sequential here, no parallel-worker conflict issue arises for this particular task.
0 comments
Sign in to comment or report what this prompt did for you.
Sign inNo comments yet. Run the prompt and report what you got.