Anthropic engineers describe shipping Claude with Claude
Original title: How Anthropic ships Claude with Claude: Separating the Brain from the Hands
Opus 5.5 is described as cheaper to run than its predecessor, with thinking blocks kept across turns to hinder distillation.
A Latent Space interview with Anthropic's Katelyn Lesse and Angela Jiang covers the Claude Opus 5.5 launch and API changes. Managed Agents keeps the session and harness separate from a sandbox that executes code, so a failed sandbox does not end the session and service credentials need not sit where generated code runs. The guests say near-term attention has shifted toward safety, and later platform ideas should be treated as plans rather than shipped features.
Unconfirmed: The S-1 profit figures and ranking of Opus 5.5 as the top model are the hosts' framing, not independently verified here.
- Early-access customers evaluate the model in their own harnesses.
- An earlier design sandboxed the whole coding agent and added startup latency.
- Dreaming is one agent inspecting another's work and writing memories.
- Hosts cite an S-1 swing from a 126% loss to 19% GAAP profit.
- Products
- ClaudeManaged Agents
- Organizations
- Anthropic
- AI models
- Claude Opus 5.5
From the original article
15 min read at latent.space
Katelyn Lesse and Angela Jiang on Opus 5. 5, the Claude Platform, and the engineering behind AI agents.
By swyx .ai · latent.space
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.