Key topics
- Planning vs execution separation and explicit step criteria.
- Tool-calling with JSON schemas and idempotency controls.
- Verification: structural checks, business rules, and grounding checks.
- Memory boundaries and retention policies.
- Human escalation triggers and approval steps.
Common pitfalls
- Letting the model invent tool parameters without validation.
- Write actions without approvals or commit evidence.
- Unbounded context and memory leading to drift and leakage.
- No failure handling: retries cause duplicates and inconsistent state.
Recommended practices
- Expose minimal tools for each role and workflow.
- Make side effects explicit and replay-safe.
- Use structured outputs for any action-ready result.
- Instrument and evaluate workflows continuously.
This page is intended to be actionable for engineering teams. For platform-specific details, cross-reference /platform/agents, /platform/orchestration, and /platform/knowledge.