Beyond chat completion
- Planning — breaking a goal into steps
- Tool use — APIs, databases, browsers, code execution
- Memory — conversation state, retrieved documents, user preferences
- Orchestration — routing work between specialized sub-agents
Patterns that work in production
1. Single-responsibility agents
2. Tool-first design
3. Human-in-the-loop for high impact
4. Observability from day one
Stack I reach for
| Layer | Examples |
|---|---|
| Orchestration | LangGraph, n8n, custom Node/Python workers |
| Models | OpenAI, Claude, local Ollama for dev |
| Retrieval | Supabase pgvector, embeddings + RAG pipelines |
| Deployment | Vercel, Railway, Docker for long-running workers |
Common failure modes
- Over-autonomy — agent sends emails or charges cards without confirmation
- Context stuffing — entire knowledge bases pasted into one prompt instead of retrieval
- No evaluation — shipping prompt changes without regression tests on real tasks
- Monolithic agents — one agent tries to research, code, deploy, and document
