Skill · doctrine · framework-choice
Pick a framework
Pick the thinnest library that supplies a missing primitive (checkpoints, types, harness, MCP, documents, handoffs). Use when choosing LangGraph, CrewAI, OpenAI Agents SDK, Mastra, Pydantic AI, LlamaIndex, OpenHands, AutoGen, or the user is shopping by stars / asking LangGraph vs CrewAI.
GET /api/canon/skills/framework-choice?format=md
Stars are a floor, not a ranking. Contribution-free clones with disabled PRs exist. One orchestrator. Libraries for missing primitives only.
When
The raw loop works and you still need one primitive. Not before evals. Not because a README went viral.
Do
- 01
Name the missing primitive
Checkpoints → LangGraph or Mastra. Types → Pydantic AI. Handoffs → OpenAI Agents SDK. Code-act → smolagents+E2B. Documents → LlamaIndex. Coding/research product → a harness (Codex, OpenHands, Cline, Deep Agents). If none, stay on the raw loop.
- 02
Keep ownership
You can still print the prompt, pause the run, and swap the model. If not, you bought a box.
- 03
One orchestrator
Do not mix LangGraph plus CrewAI plus AutoGen in one process.
Don't
- Pick the viral harness of the week.
- Rewrite the company around a framework before the golden set exists.
- Run two prompt stacks and two retry policies.
Hard rules
- Evals first. Raw loop second. Framework last.
- Coding work starts from a harness, not a blank graph.
- Read the atlas contribution, not the star count.
Refuse
- Star-driven selection — Picking the viral harness of the week. Contribution-free clones with disabled PRs.
- Two orchestration frameworks — LangGraph plus CrewAI plus AutoGen in one process. Nobody owns the loop.
- Unowned prompt — The policy lives in a framework black box you cannot diff, test, or revert.
Load with this
Load next
- Workflows versus agentsYou are about to pick a topology. Most production systems that work are workflows with a small agentic island.
- Design a harnessThe product is a desk, not a chat wrapper. Claude Code, Codex, OpenHands, Cline, Goose, Deep Agents converge here.
- Eval harnessYou need to know whether a prompt, a tool, or a skill got worse.
Trigger tests
Should fire
- “LangGraph vs CrewAI vs Agents SDK?”
- “Which framework should we use?”
- “Is AutoGen production ready?”
Should not
- “Which frontend framework?”
- “Should we use Redis?”