Descriptions:
At AI Dev 26, Melissa Herrera — Senior Developer Advocate at Temporal — tackles one of the most overlooked failure modes in production AI agents: the complete loss of context and progress when an agent crashes mid-task. Her talk introduces the concept of “durable agents” and explains why the naive fix — simply restarting the agent — is more costly than it appears.
Herrera walks through three concrete penalties of non-durable restarts: wasted tokens already burned on prior LLM calls, non-determinism (the rerun may take a completely different execution path), and lost wall-clock time in long-running workflows. She then shows how Temporal addresses all three by acting as a persistence and orchestration layer that checkpoints agent state at each step, allowing workflows to resume from the exact point of failure rather than from scratch.
The technical portion includes Python code using Temporal’s SDK, demonstrating how to wrap an agent’s steps with `@workflow.defn` and `@workflow.run` decorators and isolate failure-prone operations (like web search or LLM calls) into retryable Temporal activities. Herrera draws an explicit parallel between AI agents and classic distributed systems — arguing that agentic loops are simply distributed systems “in a different font” — and notes that Temporal supports not just agent workflows but also MCP servers, inference processes, and RAG ingestion pipelines. The talk is practical and directly applicable to teams running long-horizon agents in production.
📺 Source: DeepLearningAI · Published May 21, 2026
🏷️ Format: Deep Dive
