Durable Agents: long running AI workflows in a flakey world
summary
Shows a concrete system: Pydantic AI plus a Temporal workflow wraps every LLM or tool call as an activity so runs survive node crashes and network flakiness. Demo runs two agents playing 20 questions with 10% simulated failures, Temporal replays cached activity results in ~11ms replay instead of redoing inference, and Logfire traces token counts and costs. Supports Temporal, Deboss, Prefect, and a gateway for one-key model switching plus type-safe structured outputs validated end-to-end.
AI-written summary · about
chapters (9)
- 0:00Introduction and background
- 3:02Pyantic AI agent framework
- 8:57Why AI needs engineering
- 11:27Durable execution basics
- 13:5520 questions demo setup
- 17:42Running the workflow
- 20:12Recovering from failures
- 22:00Engineering principles for AI
- 23:05Q&A and closing
related talks
33mFrom Prototype to Production: Making Agentic AI Systems Reliable
25mAgents: Building Production AI Agents with Temporal
45mTaming AI Agents With Output.Ai
01h 00mReliable Agentic Systems need Durable Execution 🤖
26mBeyond Sandboxes: Architecting Durable Runtimes for AI Agents
56mOpening Keynote
01h 04mBuilding AI Applications the Pydantic Way
30mGenerative AI Monitoring with PydanticAI and Logfire