Long-Running Claude

Most discussion of agents has, until recently, assumed second-scale interactions: a user prompts, the agent responds, the loop closes. Anthropic's "Long-Running Claude" is part of a documented shift to hour- and day-scale autonomy: an agent that takes on a goal, plans toward it, executes across many tools, and reports back later.

That shift breaks most of the assumptions built into existing workflows. Permission models that worked for a single tool call don't work for a hundred. Logging that worked for a single response doesn't work for a multi-hour trace. The human reviewing the work needs to do so without re-doing it. "Long-Running Claude" is the lab itself naming the shift; the operating-model implications are downstream.

A second-scale agent fits inside an existing reporting structure — someone owns it, supervises it, escalates it. An hour-scale agent doesn't. The supervisor can't watch it in real time. The accountability surface is fundamentally different.

Most organizations haven't thought about who, exactly, is responsible for an agent that ran for six hours overnight and produced an outcome the human reviewer disagrees with the next morning. Whose decision was it? What does 'approval' even mean when the human is reviewing a compressed summary of decisions the agent already executed? These are organizational questions, not technical ones, and they have to be answered before the agent ships — not after the first incident.

We don't deploy long-running agents at full autonomy on day one. The progression we use with clients is staged: read-only autonomous (the agent does the work but doesn't write), gated-write (the agent proposes, the human approves), bounded-write (the agent writes within a defined scope), then autonomous-with-summary (the agent writes and reports). Each stage builds the org's trust before the next one is unlocked.

Skipping stages is the most common cause of a team rolling back an agent deployment in month two. The deployment didn't fail technically — it failed because the organization wasn't ready to trust it at the autonomy level it had been granted. The technical problem is easy. The org problem is the harder one and the one the lab post mostly leaves to the operator to solve.