Playbook
Agent Workflows
Tool use without chaos.
Agents are reliable when planning, routing, execution, and validation are explicit.
Planning
- Break requests into steps before tool calls
- Limit step count to prevent runaway plans
- Capture intermediate reasoning in logs
Tool routing
- Allowlist tools per role and environment
- Validate inputs against schemas
- Use human approval for destructive actions
Execution safety
- Timeouts and explicit failure states
- Rate limits and cost guards
- Deterministic retries only when safe
Validation
- Confirm outputs before responding
- Require citations or tool logs for decisions
- Reject outputs that violate constraints
Checklist
- Tool schema registry
- Execution audit trail
- Validation step before final response
- Failure mode test set