What is the multi-agent approach?
You combine the strengths of different agents in a pipeline. Each one does what it does best. You are the orchestrator.
Scenario 1: Building a test framework
| Phase | Agent | What it does |
|---|---|---|
| Architecture | Claude Opus | Framework spec, CLAUDE.md, docs/ |
| Implementation | Claude Code / Codex | Code, tests, Dockerfile |
| Review | GitHub Copilot | PR review, security scan |
| Documentation | Rovo | Confluence, linking to Jira |
Scenario 2: Daily QA workflow
Morning: OpenClaw collects monitoring alerts, sends a summary to Slack at 7:00 AM.
Mid-morning: Rovo in Jira assists with triage: “This bug is similar to PROJ-1234 from Q1.”
During work: Claude Sonnet / ChatGPT for quick test cases and bug descriptions. Copilot in IDE for writing Playwright tests.
End of day: Claude Opus analyzes regression results. ChatGPT writes a report for the PO.
Scenario 3: JUnit 4 → 5 migration
- Claude Opus — codebase analysis: how many tests, which runners, @Rule, custom annotations
- Claude Opus — step-by-step migration plan with rollback strategy
- Codex / Claude Code — migration module by module, running tests
- Codex — CI/CD update (Jenkinsfile/GitHub Actions)
- GitHub Copilot — review of migration PRs
- ChatGPT + Rovo — migration guide + publication in Confluence
Scenario 4: Incident response
Min 0: OpenClaw → Slack: "500 errors spike on /api/payments"
Min 1: Claude Opus (CoT) → log analysis + last deploy
→ hypothesis: PR #892 changed the card_token format
Min 5: Codex → revert PR #892, hotfix PR
Min 10: Copilot → regression test + full test suite
Min 15: ChatGPT (CO-STAR) → post-mortem
Min 20: Rovo → Jira ticket + Confluence post-mortemCheat sheet
| Need | Agent #1 | Agent #2 | Framework |
|---|---|---|---|
| Test cases from user story | ChatGPT / Sonnet | — | CRISPE + Few-Shot |
| Edge cases | Claude Opus | — | ToT |
| Bug description | ChatGPT | Rovo | ARC |
| Automated tests | Claude Code / Copilot | — | RAG + TDS |
| Business case | Claude Opus | ChatGPT | SCQA |
| Test framework | Opus → Code/Codex | Copilot → Rovo | P2P2 → TDS |
| Daily report | ChatGPT / Sonnet | — | CO-STAR |
| Incident response | Opus + Codex | Copilot + Rovo | CoT + CO-STAR |
In the next and final post: CLAUDE.md, AGENTS.md, and meta-prompting — how to configure an agent and how to write prompts automatically.