Anthropic announced on May 6, 2026 at its Code with Claude developer conference that Claude Managed Agents now ship with Dreaming, a scheduled background process that lets agents review past sessions, extract patterns, and curate their own long-term memory. The launch arrives with three other major upgrades — Outcomes, multi-agent orchestration, and webhooks — that move Anthropic's enterprise platform decisively into the agentic era.
Key Highlights
- Dreaming is a scheduled process that reviews agent sessions and memory stores, extracts patterns, and curates memories so agents self-improve over time. It is available as a research preview.
- Outcomes, multi-agent orchestration, and webhooks are now in public beta as part of Managed Agents.
- Legal AI platform Harvey reported completion rates jumping roughly 6x after enabling Dreaming in internal tests.
- Outcomes lifted task success rates by up to 10 percentage points versus standard prompting in Anthropic's benchmarks.
How Dreaming Works
According to Anthropic, Dreaming runs as a scheduled job that "reviews your agent sessions and memory stores, extracts patterns, and curates memories so your agents improve over time." The process surfaces recurring patterns that a single agent cannot see on its own, including repeated mistakes, workflows that converge across runs, and preferences shared by a team.
Crucially, Dreaming does not overwrite the original memory store. Instead, the system creates a separate revised version that developers can keep, edit, or discard, giving teams a control surface over what their agents learn. Operators can also let Dreaming update memory automatically for fully autonomous self-improvement.
In tests with legal AI platform Harvey, completion rates rose roughly six-fold purely from agents carrying institutional knowledge across sessions, according to The New Stack and SiliconANGLE coverage of the announcement.
Outcomes: Rubric-Driven Quality
Outcomes flips the agent loop from prompts to graded results. Developers write a rubric describing what success looks like, the agent produces a result, and a separate grader evaluates the output in its own context window — so the grader is not influenced by the agent's reasoning. When the result misses the bar, the grader pinpoints what needs to change and the agent takes another pass.
Anthropic reports that Outcomes improved task success rates by up to 10 percentage points compared to standard prompting, with internal benchmarks showing gains of 8.4 percent on docx files and 10.1 percent on pptx files.
Multi-Agent Orchestration and Webhooks
Multi-agent orchestration lets a lead agent break a job into pieces and delegate each one to a specialist subagent with its own model, prompt, and tools. Subagents work in parallel on a shared filesystem and contribute their findings back into the lead agent's context. Anthropic cited Netflix's platform team, which built an analysis agent that uses orchestration to inspect hundreds of builds in parallel and surface only the patterns worth acting on.
Webhooks complete the asynchronous loop: developers can define an outcome, fire off the agent, and get a callback when the work is done — eliminating the need to keep a session open while long-running jobs execute.
Why It Matters
The announcement signals a structural shift in how AI agents are built and deployed. Instead of stateless chatbots that reset between sessions, Managed Agents now learn across runs, coordinate as teams, and report back asynchronously. For enterprise teams, this closes the gap between prototype agents that demo well and production agents that get measurably better over time.
The compute-hungry workload behind these features may also explain another announcement made at the same conference: Anthropic doubled Claude Code usage limits on Pro and Max plans and disclosed a new compute deal with SpaceX targeting the Memphis data center.
What's Next
Dreaming remains in research preview and Anthropic has not committed to a general-availability date. Outcomes, multi-agent orchestration, and webhooks are open to all Managed Agents customers in public beta starting today. Developers can read the full feature breakdown in the Claude documentation and the Anthropic engineering blog.