Anthropic on shipping Opus 5.5: separating the agent's brain from its hands
Latent Space · rss · 2026-10-28
Anthropic heads of Engineering and Product, Katelyn Lesse and Angela Jiang, join Latent Space to unpack how the company ships Claude — against the backdrop of its S-1 filing showing a swing from -126% loss to 19% GAAP profit.
Key points
- Launch is also an API launch: Opus 5.5 targets intelligence-per-dollar and less verbosity; preserving thinking blocks across turns helps protect reasoning from distillation but can break customer harnesses.
- Brain vs. hands: Managed Agents decouples the session harness from the code-executing sandbox — a failed sandbox no longer kills the agent session, and environments spin up on demand.
- Security by architecture: sandboxes never hold service credentials; secrets live separately and attach via proxy with network egress controls.
- Memory experiments: "Dreaming" is simply one agent inspecting another's work and writing back memories, built on the same managed-agent primitives.
- Priorities: with launches getting harder, more attention is shifting to safety; CPU capacity has become a strategic constraint; the Messages API is being rebuilt in Rust.
More from coding & agent
- RateSpot MCP Server Lets AI Assistants Fetch Real-Time Mortgage Rates — modelcontextprotocol · 2026-10-28
- Cyphrex Launches MCP Gateway to Block AI Agent Actions with Audit Trail — modelcontextprotocol · 2026-10-28
- Reflection AI runs 1.3B sandboxes in largest RL run, peaking at 883K vCPUs — zacharynado · 2026-10-28
- Claude turns 1,256 competitor ads into a creative brief — only 349 survive 30 days — aryanXmahajan · 2026-10-28
- Kinex physical coding agent self-evolves to solve all 365 RoboCasa tasks in 7 rounds — daibond_alpha · 2026-10-28
- Free Claude Code skill repo adds 50 video effects via one /effects command — aitrendz_xyz · 2026-10-28