AI Agents & Orchestration
Tool-using LLMs that plan, execute, and verify.
When to Use
- Multi-step workflows (ETL, research, ticket triage).
- APIs/tools required (SQL, web, email, Slack, Jira).
- Stateful tasks & human-in-the-loop approvals.
Design Points
- Planner vs Worker agents; shared memory & scratchpads.
- Deterministic subroutines for critical steps.
- Safety: tool whitelists, rate limits, audit logs.
Minimal Orchestration Flow
- Classify intent → route to workflow.
- Retrieve context (RAG/DB) and constraints.
- Plan steps → execute tools (with retries/backoff).
- Verify results → produce final output.
Frameworks
- LangGraph (graph state + checkpointers)
- CrewAI / AutoGen (multi-agent patterns)
- Custom Airflow/Dagster operators for mixed jobs