Agent frameworks — LangGraph, AutoGen, CrewAI

Three agent frameworks worth knowing in 2026:

- LangGraph — explicit graph of nodes + edges. State is shared dict. Best when you want control.

- AutoGen — conversational agents with role-based prompts. Microsoft. Better for research-style multi-agent.

- CrewAI — agents as 'roles' with goals + backstory. Highest abstraction. Fastest to demo, hardest to debug.

Plus the SDKs: OpenAI Assistants API, Anthropic agent SDK. Lower abstraction, native to provider.

LangGraph models an agent as a state graph:

- State — typed dict (TypedDict / Pydantic) shared by all nodes
- Nodes — functions that read state, do work, return updates
- Edges — conditional routing between nodes

Why it wins:
- You see the graph; bugs are findable
- State is centralised; no implicit shared memory
- Persistence + checkpointing built-in
- Clean async support

Weakness: more boilerplate vs CrewAI for simple cases. Worth it past prototype.

Define a state. Add nodes. Add edges. Compile. Invoke. Same shape every time.

This 25-line example shows the structure. Real agents add tools, conditional edges, checkpointing — but the skeleton stays the same.

An agent without traces is unfixable. You'll spend hours guessing why it gave a wrong answer.

My default stack:
- LangSmith (managed) or Langfuse (self-host) — captures every prompt, tool call, latency
- One-line setup with most frameworks
- Search by user, query, error

The minute you ship an agent, ship its trace dashboard. Cost: ~$0 for low volume. Saves entire debugging cycles.

Day 75 done.

- LangGraph for control
- AutoGen for research
- CrewAI for fast demos
- Always add tracing

Tomorrow (Day 76): memory in agents. Short-term, long-term, and the patterns that survive contact with users.