Long-Running Agent

Most chatbot exchanges finish in seconds. A long-running agent might spend an afternoon migrating a codebase to a new library, or a week watching a supplier’s site and filing a report each time prices change. Along the way it waits: for a slow API, for a test suite, for a person to approve a step. The longer a run lasts, the more chances something has to fail, and the less acceptable it is to start over from scratch.

That changes what the software around the model has to do. A long task can outgrow one context window, so the agent needs notes, files or other agent memory to carry progress forward. Anthropic’s write-up on harnesses for long-running agents compares it to engineers working in shifts: each new session starts with no memory of the last, so the agent leaves a progress log and commits its work before stopping. Failures need retries that don’t repeat finished steps. Cost grows too, because every extra step in the agent loop uses more tokens. METR’s research tracks a related question from the other side: how long a task, measured in human working time, today’s agents can finish reliably.

On Chain of Thought, Oracle’s Richmond Alake frames run length as a design target in episode 53 at 42:29: if success means a coding agent can run for hours, that choice shapes which engineering techniques you need. Galileo sponsored episode 53. In episode 67 at 11:38, Cisco’s Jeetu Patel argues that price per token could become a vanity metric because long-running agents use more tokens; what matters is the cost per unit of outcome. Episode 67 is delivered by Svix, a Season 4 presenting sponsor; see the Svix read.

Disclosure: the term also appears in sponsor copy. Inngest, a Season 4 presenting sponsor that sells durable execution for agents, uses “long-running agents” in its episode 76 sponsor read. That read is advertising, not a source for this definition.