Anthropic's long-horizon agents put the verifier in its own context window

Martin's talk lays out the architecture behind Anthropic's newer managed-agents surface for long-running work. The harness running the agent loop is decoupled from the containers executing its work and connected through an append-only session log, so a crashed container does not lose the session and credentials never sit inside the sandbox itself. Verification also runs in a separate context window from the one that did the work: when the same context both produces and grades an answer, Martin described the results as unreliable and prone to fabricated self-assessment. Splitting build and verify into a loop, tested on OpenAI's parameter golf ML-research benchmark, produced markedly stronger results from frontier-capacity models. On memory, Martin described an offline consolidation pass, modeled loosely on how the brain moves short-term memory into long-term storage during sleep, that corrects errors an agent wrote to its own memory during a task. In a repeated test using an agent playing Pokemon, an uncorrected memory error caused the agent to fall through the same in-game trap door in five separate runs; runs that went through the consolidation pass avoided the error.

Why it matters: Teams could add session logs decoupled from execution and verification kept out of the working context without waiting on new model capability. The memory-consolidation piece is earlier-stage, but it targets a failure mode most long-running harnesses do not yet check: an agent's own written memory going wrong and staying wrong.