When a Claude agent "dreams," a second Claude agent does the dreaming. Angela Jiang, who works on Anthropic's Claude Platform team, described the mechanism on the Latent Space podcast. The dreaming feature in Claude Managed Agents starts another managed agent. That agent inspects what an earlier agent did and writes memories back for it. "So it's managed agents all the way down," Jiang said.
Jiang and her colleague Katelyn Lesse appeared in an episode published October 27, 2026. Much of it covered how Anthropic launches models safely, including the security design that separates an agent's "brain" from its "hands." A smaller part of the conversation was aimed at developers deciding whether to build their own agent machinery or use Anthropic's. Jiang talked about how the managed version is built, what Anthropic plans to open up next and the two architecture rules she thinks matter for multi-agent systems. The conversation also covered evals, the tests developers use to check whether an agent does its job.
What managed agents are
An AI agent is a language model that works through a task in steps: it calls tools, reads the results and decides what to do next. The loop that runs those steps is called a harness. The agent also needs somewhere safe to run code, usually an isolated environment called a sandbox.
According to the discussion, Anthropic's Messages API is essentially raw access to the model: a developer sends text, gets the model's output back and has to build the rest. Claude Managed Agents is a set of APIs layered on top. A developer defines an agent, starts sessions with it and skips setting up the harness and infrastructure. Anthropic's documentation says developers choose the model, instructions, tools, skills and connections to outside services, while Anthropic supplies the agent loop, tool execution and the runtime. Sessions keep their history and outputs on Anthropic's servers, and an application can stream an agent's progress, steer it or interrupt it.
The conversation noted that many developers barely looked at the product when it launched because they were busy building their own harnesses. The managed service also shares a good deal of plumbing with Anthropic's own products. The platform team runs the sandboxing fleet behind both Managed Agents and products such as Claude Tag. The two products don't share a harness, because each needs different behavior at that layer.
Safety work has slowed the feature work
Asked whether the memory work from earlier in the year had stalled, Jiang said the team has slowed down on skills, dreaming and memory compared with earlier in the year. She quickly added that "we love skills" and that updates are still coming. The reason given on the show was that launching models has become harder from a safety perspective. Less of the team's energy goes into features that push up an agent's performance and more goes into building safety into the agent system.
Jiang said the harness Anthropic provides is "intentionally simple." The team does not want to be overly opinionated, she said, and wants to make sure best practices get merged in. She described some pieces of the agent loop as "undifferentiated": hard to string together, yet the kind of boilerplate everyone needs. The more interesting work, in her account, is the strategies Anthropic builds on top of that base layer to "hill climb," the engineers' term for improving a system step by step against a measure.
Dreaming, and agents checking agents
Dreaming is one of those strategies. Anthropic introduced it in a May 19, 2026 update. That update separates memory an agent writes while it works from dreaming, which reviews past sessions afterward to collect recurring lessons, clean up stale notes and record shared preferences. The Dreams documentation describes it as an asynchronous job over an existing memory store and between one and 100 earlier sessions. It merges duplicates, updates entries that turned out wrong or stale and writes the result to a separate memory store, so developers can compare it with the original or reject it.
Jiang said dreaming is "a multi-agent system that we hill climb on top of" Managed Agents, so the product is running on itself. That was deliberate, she said, because the team expected developers to combine modular pieces in interesting ways. She listed other patterns: best-of-N, where several attempts are run and the best one is kept, and a verifier agent tied to a goal or "outcome." The same May update describes Anthropic's outcomes feature as a separate grader, with its own context window, that judges a deliverable against a rubric and sends feedback for another attempt.
Developers have gone further on their own. Jiang described seeing "some pretty funky, crazy stuff," including a harness that shares memory across everything and starts large numbers of parallel agents. Her response was that Anthropic should let developers build such systems the way it built dreaming. She said an update will come "relatively soon" that exposes the underlying building block, so users can construct their own multi-agent strategies, with templates showing how dreaming was built as a starting point. She gave no date. Managed Agents can already coordinate several agents: the May update describes a lead agent handing work to specialist agents that run in parallel on a shared filesystem. The planned change would let developers design those coordination strategies themselves.
Jiang's two rules for multi-agent setups
At the end of the episode, Jiang's call to action for listeners was to try a multi-agent setup. "I think you will find it to be very good," she said. When pressed for specifics, she gave a short recipe. Generally you want one lead agent that splits work into parts. If the task can be done in parallel and the agents share the same context, she expects "a lot of gains," though she said it depends on the problem.
After that, she said, there are only two things to care about in the architecture. The agents need a shared state, and the sub-agents need to run asynchronously so they don't block one another. "Other than that, like, let it rip and, like, see what Claude designs for you," Jiang said. She expects Claude to get better at working out how to "split itself, clone itself" and fit the architecture to the problem. She also said she wouldn't be surprised if a setup like this could crush a problem people had found intractable, "in maybe one shot or something."
The conversation noted that agents can now assign roles among themselves. Developers could always give agents jobs, but now the agents can give themselves jobs. Jiang agreed.
Most teams need more evals
The evals advice in the discussion was blunt. There are many kinds of evals, from fixed, deterministic checks to using a language model as a judge. Most teams need more of them rather than fewer, because no small set will answer every question about how to improve a system. Under-investing in evals was described as perhaps the biggest mistake Anthropic has seen. Anthropic's guide to agent evals draws on a similar mix. It combines code-based checks, model-based judgments and human review, and warns that model judges need calibrating against people.
Jiang added that the first evals can be simple. She described teams that study the whole menu of eval techniques, trying to work out which is best, while the number of evals they actually have is zero. Doing something simple, she said, "will help."
The discussion then turned to treating engineering like research. Under this approach, the team building a harness or product defines its goal up front, turns it into an eval and then asks why the system fails. Sometimes the eval is wrong, sometimes the environment is wrong and sometimes the model is. When the point about faulty evals came up, Jiang answered "Totally." When the conversation turned to application engineering and model research borrowing each other's methods, she called it "a whole new skill set."
Anthropic is likely to offer more help here, according to the discussion. A product for creating evals for managed agents was described as a probable addition, and using it would feel much like working with Claude, which was described as really good at creating evals. The conversation also covered presentation. People's attention fades when they stare at raw data files, and a clean interface for comparing two outputs side by side lets developers do the work that hired data labelers would otherwise do. Claude can then take the resulting ratings as feedback and improve its own work. Jiang called that "a very, like, AGI-pilled way" to work, because it lets developers update the system methodically as models improve.
A worked example from claude.ai
The conversation pointed to a recent Anthropic engineering post as an example of goal-driven work: an account of making claude.ai faster, published September 23, 2026. On the show it was recalled as nine of 12 goals reached in about three days, against a two-week deadline. The post gives different figures. During a two-week sprint in August, the team turned four user journeys covering 95% of activity into 13 measurements, and 12 of the 13 initial targets were met by day three. For real users at the 75th percentile, the time until a fresh claude.ai page could accept typing fell from 3.1 seconds to 0.55 seconds, and the team reports an overall improvement of about three times.
Jiang said she loves it when Anthropic shares its actual prompts, because people expect some unusual technique. Sometimes, she said, the prompt is just: "try harder, Claude."