Alex Poliakov

Talks

• Building Crashproof Agents

  Beyond the problems of accuracy and cost, AI agents are challenged by long-running execution processes, human-in-the-loop approval logic, non-deterministic call paths, and upgrade complexity. Agents are expected to perform more important tasks over time, with increasing costs of failure. These challenges can be met with Durable Execution, a set of techniques that originated in transactional processing and is now growing in popularity with agent builders. This talk introduces durable execution and the guarantees it can, and cannot, provide, then shows why durability is becoming a key pattern in AI agents. It closes with how DBOS achieves durability using only Postgres, including advanced maneuvers like queueing, upgrading long-running workflows, and forking. For developers with experience in agentic development or an interest in reliable software. Postgres experience is a plus.