Sub-agents are great when your main agent needs parallel help (research, drafting, coding, data cleanup). But many users hit the same problem: they spawn extra agents and lose track of who did what.

When to Use a Sub-Agent

• Use sub-agents for large tasks that can be split (research + writing + QA).
• Do not use sub-agents for tiny one-step tasks (you'll add overhead for no gain).
• Always keep one manager (your main agent) responsible for the final merged result.

2-Minute Setup Pattern (Manager + Workers)

Copy/Paste Prompt Template

You are the manager agent.
Create sub-agents only if needed.

Goal: [one clear result]
Deadline: [time]

Assign workers:
1) Research worker: gather facts/sources only
2) Draft worker: produce first draft using research
3) QA worker: find errors, contradictions, missing steps

Rules:
- No overlapping responsibilities
- Each worker must return output in this format:
  - Task completed
  - Key findings (3-7 bullets)
  - Open questions
- After workers finish, merge into ONE final answer
- Final answer must include:
  1) Executive summary
  2) Action steps
  3) Risks/unknowns
  4) Next best action

High-Friction Mistakes to Avoid

1) Letting workers answer the user directly

Have workers report to the manager only. The manager sends the final response.

2) Assigning overlapping tasks

If two workers both "research" or both "draft," you get duplicates and contradictions.

3) No merge rule

Without a forced final merge, you end up with fragments instead of a deliverable.

4) No evidence requirement

For factual work, require each worker to cite source links or data origin.

Advanced Troubleshooting (Real-World Failure Modes)

1) “The workers finished, but I can’t find the output”

This is usually a destination mismatch (different thread/session/surface), not failed work.

• Confirm where the manager is supposed to post final output (same thread vs another channel).
• Require the manager to echo: destination + session/thread id + timestamp before final delivery.
• If using scheduled/background runs, verify delivery target explicitly.

2) “I keep getting duplicate final answers”

You likely have multiple managers or stale workers still active.

• Enforce exactly one manager role.
• Ask the manager to list active workers, then stop duplicates.
• Add a strict rule: only manager can publish final answer.

3) “Workers time out or return partial work”

Large tasks fail when each worker is too broad.

• Split work into smaller chunks with explicit limits (max 5 sources, max 10 bullets).
• Require each worker to return a blocked reason if incomplete.
• Have manager run a gap-check pass before merging.

4) “Manager says merged, but key details vanished”

This is a merge-quality issue.

• Force a merge checklist: dedupe, conflict resolution, missing-step scan, citation carry-over.
• Require manager to include a short what changed during merge section.

5) “Can I use a sub-agent to keep files organized?”

Yes — but only with strict file rules, or you'll create duplicates/triplicates fast.

• Create exactly one File Steward worker role. No other worker can create/move files.
• Enforce find-before-create: worker must search for an existing canonical file before writing.
• Require a completion proof block: file path + action + reason + timestamp.
• Manager must approve any new top-level folders before the worker creates them.
• If file chaos already exists, pause new writes and run a one-time dedupe pass first.

If It Still Feels Messy

Run this reset workflow:

1. Stop creating new workers.
2. Ask the manager to list all active workers and their exact jobs.
3. Kill overlapping workers and keep only distinct roles.
4. Re-run with strict format + final merge required.