Beyond Sandboxes: Architecting Durable Runtimes for AI Agents
summary
Argues production AI agents need an orchestration layer that persists state, checkpoints every step, and recovers from crashes — demonstrated by killing a run that had scheduled 10 tool calls (8 succeeded, 2 stuck) and restarting it to complete. Uses open source Conductor plus an AgentSpan runtime that compiles agent loops into durable workflows, records LLM calls (Anthropic/Sonnet referenced), tool executions, retries, human-in-loop pauses, and full audit trails via Postgres/Redis. Key ideas: late-bound sagas, checkpointed tool calls.
AI-written summary · about
chapters (7)
- 0:00Durable agent runtimes intro
- 3:19Sandboxes necessity and limits
- 8:36Runtime vs orchestration
- 13:33Conductor as durable substrate
- 20:06Debugging agent executions
- 22:55Scaling agent workflows
- 24:18Replayability and context
related talks
33mFrom Prototype to Production: Making Agentic AI Systems Reliable
36mHow to Safely Deploy AI Agents that Write and Manage your IaC
30mArchitecting Integrity: Responsible Software Evolution in the AI Era
20mBuilding Deterministic Infrastructure for Non Deterministic AI Agents
44mBuilding Agentic AI Systems: From Transactions To Conversations
19mNo Token Budget: Agentic Software Factories on Kubernetes
26mbuilding reliable AI systems in production
24mDurable Agents: long running AI workflows in a flakey world