Dealing with versioning in long-running agents

The blog post discusses the challenges and solutions for versioning in long-running AI agents, emphasizing the risk of silent failures when agents are updated mid-execution. It highlights the problem with traditional versioning approaches, which struggle with the unpredictability and long execution times of AI agents, resulting in potential mismatches in execution history interpretation. The post introduces a solution using Restate, a system that enforces immutable deployments and pins each execution to the version it started with, thus preventing mid-execution version mismatches. It describes how Restate uses durable execution to ensure consistent processing by recording and replaying every non-deterministic step in an agent's operation, and offers escape hatches for situations where immutability is too rigid. The infrastructure-focused approach shifts versioning concerns away from application code, providing a more reliable and auditable framework for handling AI agent executions across different versions.