Claude for Long-Horizon Tasks — Lance Martin, Anthropic

A long-running agent is easier to recover when its work history lives outside the process doing the work. In this architecture discussion, an engineer at Anthropic describes separating the agent's controlling process, its execution containers and a durable session log.

The log records events by adding new entries while retaining earlier ones. If the controller or a work container fails, the recorded session survives separately. The agent can also retrieve older material from that log instead of depending entirely on a shortened conversation summary. This makes the work history both recoverable and available for later inspection, rather than tying its survival to a single running container. It is a recovery boundary, not a guarantee that external actions are never repeated or lost.

“giving Claude access to a bunch of your secrets and letting it run for 10 hours and not watching it can be a little bit spooky”

Lance Martin

“instead of encoding steering me and into like me as the human, you're encoding the signal into the environment”

Lance Martin

“Five out of five replicates with raw memory store fell down this trap. With the dreaming, this error is corrected”

Lance Martin

“Let the model structure and maintain its own memory. Don't give it a prescribed memory schema.”

Lance Martin

“org level harnesses are real leveler of the playing field”

Lance Martin