Technology Radar
Ignoring durability in agent workflows is an anti-pattern we’ve seen across many teams, resulting in systems that work in development but fail in production. The challenges facing distributed systems are even more pronounced when building with agents. A mindset that expects failures and recovers gracefully outpaces a reactive approach.
LLM and tool calls can fail due to network interruptions and server crashes, halting an agent's progress and leading to poor user experience and increased operational costs. Some systems can tolerate this when tasks are short-lived, but complex workflows that run for days or weeks require durability.
Fortunately, durable execution is being integrated into agent frameworks such as LangGraph and Pydantic AI. It provides stateful persistence of progress and tool calls, enabling agents to resume tasks after failures. For workflows that involve a human in the loop, durable execution can suspend progress while awaiting input. Durable computing platforms such as Temporal, Restate and Golem also provide support for agents. Built-in observability of tool execution and decision tracking makes debugging easier and improves understanding of systems in production. Teams should start with native durable execution support in their agent framework and reach for standalone platforms as workflows become more critical or complex.
