Evolution of agentic surfaces — Gagan Bhat & Isabella Kai He, Anthropic

An agent can start reasoning sooner if its reasoning loop does not have to wait for the environment that runs its tools. The description of this Anthropic session explains an architecture that separates those components, allowing reasoning to begin while the tool container starts. Anthropic's described design also uses a durable session log to resume the reasoning process after a failure, while a failed tool environment can be retried separately. This makes startup and recovery properties of the surrounding software, rather than tasks the model must solve through conversation alone.

“Sonnet 4.5 developed what Anthropic's Applied AI team came to call context anxiety: approaching its context window limit, it would wrap work up early and stop with room to spare. They built context resets into the harness to compensate. Then Opus 4.5 shipped without the behavior, and the fix turned into pure overhead, adding latency and discarding cache it should have kept.”

AI Engineer

“a harness encodes assumptions about what the model cannot do on its own, and those assumptions go stale as models improve.”

AI Engineer

“The architectural consequence is decoupling the brain, meaning the agent loop, from the hands, meaning the tool execution environment. Both started in one container, so the model could not begin reasoning until setup finished and either half failing took the whole agent down.”

AI Engineer

“Splitting them lets reasoning start while the container builds in parallel, which they measured at 60% faster time to first token at P50 and over 90% at P95.”

AI Engineer

“That log ends up doing triple duty, providing observability, letting the harness read context slices back in after Claude discards them mid run, and feeding a periodic batch process they call dreaming that rewrites the agent's memory so the next day's sessions start smarter.”

AI Engineer