Five patterns that actually coordinate agents.

What's inside

• 01 Researcher + Writer Pipeline

  3 agents

  Orchestrator assigns research to one agent, writing to a second, and editing to a third. Includes the handoff format, the revision loop prompt, and how to stop when research quality is too low to proceed.

• 02 QA Loop

  2 agents

  A producer agent plus a verifier that checks output against a criteria spec. Includes the criteria spec format, the loop termination prompt, and how to detect when the same criterion fails twice in a row, which usually means the spec is broken, not the producer.

• 03 Code Review Pipeline

  3 agents

  Author writes, Reviewer checks, Architect makes final calls on escalated design questions. Each agent has a distinct system prompt. Includes the escalation protocol and a check that prevents the Reviewer from approving code with unresolved blocking issues.

• 04 Parallel Fan-Out

  N + 1 agents

  Orchestrator sends the same question to N specialist agents simultaneously, then a Synthesizer combines the answers. Includes the synthesis prompt and conflict resolution logic: when to defer to the domain expert, when to flag an unresolved empirical conflict, and when to present a tradeoff and let the user decide.

• 05 Long-Running Task with Checkpointing

  1 agent

  Agent breaks a large task into numbered steps, saves state to a JSON checkpoint file after each step, and recovers from interruptions without re-running completed work. Includes the checkpoint file format, the recovery prompt, and what to do when a step fails partway through.

Sample: QA loop verifier system prompt

You are a verification agent. You receive output and a criteria
spec, and you check whether the output meets all criteria.

Input format:
- CRITERIA: numbered list of requirements
- OUTPUT: the content to verify

For each criterion, check: PASS or FAIL.

Output format:
---
VERIFICATION REPORT
Criterion 1: [criterion text] — PASS / FAIL
  [If FAIL: one sentence explaining what is wrong and
  what would make it pass]
Criterion 2: [criterion text] — PASS / FAIL
  [If FAIL: explanation]
(continue for all criteria)

OVERALL: PASS / FAIL
---

Rules:
- Check each criterion independently.
- Do not add criteria that are not in the spec.
- OVERALL is PASS only if every criterion passes.
- Be precise. "Criterion 3 fails because the function throws
  on null input but the spec requires it to return an empty
  array" is good. "Code has issues" is not.

Every template follows this pattern: the system prompt defines the input format, the output format, and a short rules list. The rules list is where the behavior actually lives. This verifier's rules prevent it from inventing criteria or approving output with a single failing check. The download includes complete prompts for all 5 templates plus each sub-agent role.

Who it's for

• Developers running multi-agent pipelines who keep rebuilding the same coordination logic from scratch
• Engineers who have a working single-agent setup and want to add a review or verification layer without redesigning everything
• Teams building internal tools where agents hand work to each other and need a reliable format for passing context
• Anyone who's had an agent loop indefinitely and wants a termination strategy that doesn't require watching it

Questions

• Does this work with any LLM or just Claude?
  Any LLM. The patterns are model-agnostic. The prompts use plain instructions with no Claude-specific syntax, API calls, or tool definitions. The templates work whether your agents are Claude, GPT-4, Llama, or a mix of all three.
• What format is the download?
  Markdown file and PDF. The Markdown is formatted for copy-pasting prompts directly into your agent framework or config. The PDF is formatted for reading. Both contain identical content.
• Can I customize the prompts?
  Yes, that's the point. Each template is a starting structure, not a fixed script. The prompts are plain text. Change the output format, add domain-specific rules, adjust loop limits, or drop an agent role that doesn't fit your pipeline.
• Do I need to run all agents on the same model?
  No. The templates work with mixed models. The coordination protocol is just structured text passed between agents. You can run your orchestrator on a large model and your verifiers on a smaller one. The templates note which roles benefit from a more capable model and which do not.

Pricing

About the author