AI 原生开放教材

外观

AI 工程从零到一（中文）源码研读·中文原文 ↗原件 ↓仓库 ↗

AI 工程从零到一（中文） ​

Given a multi-agent production deployment plan, recommend the durable-execution substrate.

Produce:

  1. Load profile. Concurrent agent-runs (p50, p99). Per-run duration (seconds to hours). Fraction of runs requiring human-in-the-loop waits. Deploy frequency.
  2. State profile. Size of per-run state (KB to MB). Retention requirement (seconds of checkpoint history, or full audit log). Determinism: can runs be replayed from checkpoints deterministically, or only from logs?
  3. Side-effect profile. Which side effects need exactly-once (payments, external APIs, email)? Which can tolerate at-least-once (pure tool reads)? Outbox pattern needed for exactly-once.
  4. Recommendation tier.
    • Tier 1 (Bedi's rule): FastAPI + Postgres. Under ~100 concurrent runs, sub-hour durations, simple retries.
    • Tier 2: LangGraph runtime or Temporal. Hour-long runs, interrupt/resume, structured retries.
    • Tier 3: Custom with outbox + event sourcing. Specialized needs, high throughput, strict audit.
  5. Deploy model. Single version or rainbow/canary? Rainbow required for long-running stateful workloads.
  6. Async / thread boundary. Which parts are async (LLM calls, tool I/O) and which are threads/processes (CPU-bound post-processing, embedding).
  7. Observability. Per-run traces, super-step audit, retry counter. Storage for traces (separate from checkpoint store).

Hard rejects:

  • Recommending Temporal for a 10-concurrent-run prototype. Ceremony cost > value.
  • Thread-per-job LLM call architectures. I/O-bound + 1MB/thread does not scale.
  • Designs without outbox pattern for paid side effects. Duplicate charges are expensive.
  • Single-version deploys for multi-hour agent runs. Users lose state on every code push.

Refusal rules:

  • If load is unknown and untested, recommend Tier 1 plus load testing. Premature optimization burns time.
  • If the user wants a tokenized / blockchain-persistent system, say that durable-execution engines typically do not solve that (write your own event sourcing); recommend legal review for tokenized flows.
  • If the team has no on-call engineer, Temporal / LangGraph runtime maintenance is under-provisioned; recommend Tier 1 until on-call is staffed.

Output: a two-page brief. Start with a one-sentence recommendation ("Tier 1 (FastAPI + Postgres + outbox) for current load; escalate to LangGraph runtime when p99 run duration exceeds 10 min or concurrent runs exceed 200."), then the seven sections above. End with a 90-day upgrade path: metrics to watch, threshold for escalation, runbook outline.