Opus 5.5 has spent a month as the model most people point to when they say “best LLM right now.” In a Latent Space conversation, Anthropic’s head of engineering Katelyn Lesse and head of product Angela Jiang talked through the launch. The model barely came up. They spent their time on plumbing, and the plumbing is the part worth copying.
Split the brain from the hands
In Anthropic’s Managed Agents, the session log and harness (the brain) live apart from the sandbox that runs code (the hands). The harness persists. Execution environments get created when needed. If a sandbox dies, the session survives.
Anthropic didn’t start there. By the guests’ account, an earlier design put the whole coding agent inside a sandbox behind an API. Booting the environment and the agent added work before the first token came back, and when the sandbox vanished, the running agent went with it. That design is common because it’s the obvious one: wrap everything in a container and call it safe. It is slow, and it is fragile.
The security argument is stronger still. If generated code runs in a sandbox that never receives a service credential, the code has no credential to read. Anthropic holds credentials separately and attaches them through a proxy, with controls on where sandbox traffic can go. Prompt injection can’t exfiltrate a key from a place the key never was. That beats any amount of “please don’t leak secrets” in a system prompt.
If you’re building an agent runtime, adopt this as your default. Ask three things: what has to survive a crash, what executes untrusted work, and which secrets actually need to be present where code runs. Most teams I see would answer the third question badly.
Memory from the same parts
The split also makes new features cheap. Anthropic’s “Dreaming” is one agent reading another agent’s past work and writing memories back. It needs no new subsystem, only the same managed-agent primitives combined differently. Lesse and Jiang say they want the base harness to stay simple and leave verifiers, parallel agents and similar strategies to developers. That’s the right instinct. A thin, reliable core with room on top ages better than a clever monolith.
The part I’d push back on
Not every launch change was free for customers. To make it harder to extract model reasoning for distillation, Opus 5.5 leans on preserving thinking blocks across turns. The guests acknowledged that a new requirement like this can break a customer’s existing harness.
I understand why Anthropic did it. Distillation is a real threat to a frontier lab. But this one protects Anthropic, and integrators pay for it in migration work. Developers who wrote harnesses that trim or rewrite conversation history should check theirs before upgrading. The pitch for 5.5 is lower cost (Anthropic says it’s cheaper to run than its predecessor, with savings depending on workload and cached context). A broken harness eats that saving quickly.
The verbosity complaints about earlier Opus models apparently landed too. Intelligence per dollar was the explicit target this round.
Evals look more like research now
One quieter point stuck with me. Early-access customers run the model in their own harnesses and report odd behavior, sometimes as a formal eval and sometimes as an offhand impression. A benchmark can’t show Anthropic how its model behaves inside a stranger’s agent loop, and those reports can. The guests describe harness work as research: set a goal, measure, inspect failures, iterate, with models helping build the evals.
The team says its near-term attention has moved toward safety as launches get harder, and that any roadmap items mentioned are plans, not promises. So memory and learning, the features people actually want from agents, are waiting behind that work. Nobody on the call put a date on them.