When to use this skill
- When building a multi-agent workflow based on the OpenAI Agents SDK
tool calling,handoff,guardrailWhen designing the agents needed for this- When you need a production-ready agent loop structure, not a prototype
Instructions
Step 1) Decompose the task goal into an agent topology
- Start with a basic structure of 1 Coordinator + N Specialists
- Keep each Specialist to a single responsibility (retrieval/analysis/execution/verification)
- Record handoff conditions as explicit rules, not natural language
Step 2) Build the minimal execution loop first
- The first step is a minimal loop of "one input → one handoff → one result"
- Collect tool call failures/timeouts in a standard error format
- Logs are recorded based on
agent,tool,handoff,latencycriteria
Step 3) Add Guardrails
- Explicitly state prohibited actions (destructive commands, external transmission) as policy
- Fix the output schema (JSON schema) to stabilize downstream pipelines
- Specify retry count/stop conditions (max turns)
Step 4) Quality check
- 3 types of cases: normal, boundary, failure
- Check the handoff false positive/false negative rate
- If errors repeat, fix the routing rules before the prompts
Examples
- Request: "I want to build an agent pipeline that separates research/summarization/verification"
- Application: Coordinator classifies intent, then hands off to Researcher/Writer/Verifier
- Result: Store step-by-step artifacts and verification logs separately
Best practices
- Do not start with 5 or more agents from the beginning.
- Record handoff rules as conditional expressions rather than sentences.
- Fix the tool interface to keep model replacement possible.
- Solve failure recovery with a fallback route (another specialist) rather than retries.