This is not a marketing label. It is a real product feature.
At Code with Claude SF on May 6, Anthropic announced a new capability for Claude Managed Agents called Dreaming. After an agent finishes work, it can use idle time to review past sessions, extract patterns it did not notice in the moment, and turn them into long-term memory. The next time a similar task appears, its behavior can change.
In plain terms, older AI agents often finished a task and then effectively started from zero on the next one. With Dreaming, the gaps between agent runs are no longer wasted. The agent reflects, summarizes, and updates its own memory store.
Anthropic uses a human sleep analogy for this: the agent is, in effect, reviewing the day after work.
What Dreaming actually does
Technically, Dreaming is a scheduled process, not something the agent does while actively working. It runs during idle periods, scans prior conversations and the existing memory store, and looks for three kinds of signals:
- Repeated mistakes, such as an agent failing at the same class of SQL query five times in a week, a pattern that may be invisible within any single session.
- Converged workflows, or paths that repeatedly led to successful completion of similar tasks.
- Team-level preferences, including shared naming conventions, code style, and reporting formats across multiple agents.
After extraction, the memory system is not merely appended to; it is restructured. Anthropic summarized the distinction this way: "Memory lets each agent capture what it learns as it works. Dreaming refines that memory between sessions."
The split matters. The agent learns while working, then consolidates after the work. That is what separates Dreaming from a conventional RAG setup or a larger context window.
Developers get two control modes. They can let Dreaming write new memories automatically, or require human review before each new finding takes effect. The first option fits unattended agent deployments; the second is more suitable for enterprise workflows that need auditability.
Dreaming is currently in research preview.
Outcomes gives AI the acceptance test
Anthropic also introduced Outcomes, a separate feature aimed at helping agents decide whether a task was completed well enough.
The setup is straightforward: a developer writes a rubric that defines success. The system then runs a separate grader agent in its own context. If the grader finds the result below standard, it points out the specific problem and the original agent tries again.
The key is that the grader has a separate context. This is not self-grading, which is almost guaranteed to pass. It is a clean review pass designed to avoid the bias of an agent reverse-engineering the standard it is supposed to satisfy.
Anthropic says internal tests showed Outcomes lifting task success rates by up to 10 percentage points versus a standard prompt loop, with larger gains on harder tasks. In file-generation scenarios, docx quality improved by 8.4% and pptx quality by 10.1%.
A 10-point benchmark gap is not small. It is roughly the kind of lift teams often hope to get from moving up a model generation. Here, Anthropic is getting it from a better collaboration mechanism.
Outcomes is now in public beta.
Multi-agent orchestration moves agents from solo workers to teams
The third release is Multi-agent Orchestration. It is less conceptually radical, but its practical value is direct.
The feature lets a lead agent assign work to several specialist agents. Each specialist can run a different model, prompt, and tool set, while progress remains traceable inside Claude Console.
Anthropic's example is an incident investigation. A lead agent sends four sub-agents to inspect deployment history, error logs, metrics, and support tickets in parallel. Once they finish, the lead agent combines the findings into a conclusion.
This mirrors the architecture Anthropic uses internally for code review: multiple reviewers work in parallel, and a separate verifier checks the result. The company is now packaging that pattern for customers.
Why the package is bigger than a feature launch
Seen together, the three features form a clear loop: Dreaming lets agents improve across tasks, Outcomes lets them evaluate completed work, and Multi-agent Orchestration lets them coordinate with peers.
That is a self-improvement loop: work, evaluation, review, memory update, and then more work. It can operate with less direct human intervention.
That is why Code with Claude SF mattered even without a new model announcement. Anthropic's ambition here is larger than simply shipping another Opus point release.
Benchmark gains from each new model generation are becoming harder and more expensive. Agent capability can still improve through coordination mechanisms and accumulated experience. If Dreaming works reliably in production, Anthropic has found a speed-up path that does not depend only on pushing scaling laws harder.
There is risk, too. An agent that writes its own memory can also harden the wrong lesson. Anthropic offers a human-review option, but whether enterprises will allow fully automatic memory updates is something the market will have to prove over months of use.
Will Dreaming become a standard feature for the next generation of agents? Probably. OpenAI's Memory and Google's Personal Context are already moving toward personalized memory, but Anthropic is abstracting the idea from user preference to agent work capability. If that abstraction holds, rivals will have to answer it.
Code with Claude SF also included other numbers: Mercado Libre's 23,000 engineers have a KPI for 90% autonomous code work by Q3 2026; Claude Code's five-hour limit doubled; API call volume rose 17-fold year over year.
But the most important signal was not the numbers. It was that agents are beginning to learn how to get better on their own.
Sources: Anthropic is letting Claude agents 'dream' so they don't sleep on the job (SiliconANGLE); New in Claude Managed Agents: dreaming, outcomes, and multiagent orchestration (SD Times); CocoLoop; Live blog: Code w/ Claude 2026 (Simon Willison)