Stateful workflows
步骤在调用间持久化。节点重启不会丢失进行中的订单。
Retry 预算
每步骤的 retry 计数 + 指数退避。耗尽时进入 dead-letter queue。
Replay debug
每步骤捕获输入;对已知坏输入 replay 可精确复现失败。
02 · Reference architecture
What we actually ship.
Every system we build follows this shape. Client at the edge, tools in a sandbox, traces everywhere, evaluators gating output. No black boxes, no "it works on my machine."
EDGE
Client
web · mobile · API
Gateway
auth · 限流 · PII 编辑
编排
Orchestrator
planner · router · memory
智能体 & 工具
Retrieval
混合 · rerank · cite
Tool calls
sandboxed · timeboxed
Evaluator
gates · rubrics · LLM-judge
数据 & 信任
Vector + BM25
租户隔离
Traces / logs
OTel · 可重放
签名输出
可审计 · 可回滚
04 · Run
Watch an agent do the job.
Three real production scenarios, replayed at observed latency. Every box is a span; every span has tokens, cost, and an eval gate. This is what shows up in your traces, not a marketing animation.
POST/api/v1/agent/runApp 收到 first-notice-of-loss。智能体须在 <4s 内路由、设准备金并索取材料。
trace · claimsspan_id 7c1f…
orchestrator.run
PLANorchestrator.plan
Plan rationale
FNOL：校验保单、照片视觉、欺诈信号、严重度与准备金、赔付门禁。
Subtasks
retrieval.policytool.visiontool.fraud_checkreasoner.severityevaluator
0ms/2.33s
LATENCY0msbudget 3.00s
TOKENS0in + out
COST$0.0000budget $0.025
EVAL GATE—deterministic + LLM-judge
已路由 · 准备金 $4 800
严重度 LOW。已设准备金，钣喷网络已启动，向投保人索取 2 份材料。
latency2.78scost$0.018tokens2,210evals7/7
09 · STACK
Modern tools, composed cleanly.
Runtime
Temporal, Inngest
Observability
OpenTelemetry, Langfuse
Deploy
Vercel, AWS, GCP
编排
Workflows, sagas
09 · FAQ
FAQ · Durable Execution
从 kick-off 到生产 rollout 八到十四周。如果与我们一起做 discovery,则包含;否则从您现有的 spec 开始。
可以。Durable execution 在每个 step 捕获输入。对已知坏输入的 replay 可精确重现失败,包括模型可 seed 时的非确定 LLM 调用。