AI 工程从零到一（中文）
Given a stack, latency budget, required primitives, and operational shape, pick a runtime.
Decision:
- Python + FastAPI + thousands of short-lived agents per second -> Agno.
- TypeScript + Next.js/Vercel + unified multi-provider -> Mastra.
- Durable state, explicit graph, resume-on-failure -> LangGraph (Lesson 13).
- Claude-first product, wants the Claude Code harness shape -> Claude Agent SDK (Lesson 17).
- OpenAI-first product, wants handoffs + guardrails + tracing -> OpenAI Agents SDK (Lesson 16).
- Multi-agent team, actor-model concurrency, fault isolation -> AutoGen v0.4 / Microsoft Agent Framework (Lesson 14).
- Role-based collaboration or event-driven deterministic workflows -> CrewAI Crew or Flow (Lesson 15).
- None of the above -> direct API calls + the stdlib loop from Lesson 01.
Produce:
- A short decision document: stack, latency target, primitives needed, observed trade-offs.
- A minimal scaffold in the chosen runtime.
- A migration plan if another runtime is in use today.
Hard rejects:
- Picking Agno or Mastra purely on "performance" when the workload is one slow call per request. Performance is rarely the bottleneck.
- Picking a TypeScript runtime in a Python monorepo without a rationale. Mixed-language agent code is an operational tax.
- Picking LangGraph for stateless short tasks. The checkpointer adds overhead that a simple workflow (Lesson 12) avoids.
Refusal rules:
- If the user wants "all five runtimes, to compare," refuse. Benchmark on your workload; framework vendor benchmarks are directional.
- If the user wants to self-host Mastra's
ee/features, refuse and point to the license terms. - If the product needs long-running async work (hours-to-days), refuse self-hosted and route to Claude Managed Agents or a queue-based architecture (Lesson 29).
Output: decision doc + scaffold + README. End with "what to read next" pointing to Lesson 24 (observability) and Lesson 29 (production runtimes) for the operational layer above the framework.