Anthropic announced a new capability for its Claude Managed Agents platform on May 6 that changes how AI agents handle memory and improvement. The company calls it “dreaming,” and if it works as described, it could meaningfully shift how businesses think about deploying AI agents for ongoing work.

The announcement came at Anthropic’s Code with Claude 2026 developer conference, alongside several other updates to the Managed Agents platform.

What Dreaming Actually Does

The dreaming feature is a scheduled background process that runs between active agent sessions. While your Claude agent is not actively working on a task, it reviews everything it has done before: past conversations, completed tasks, logged memory, and interaction patterns.

From that review, it does three things:

Prunes stale information. Notes and memory entries that are no longer relevant get cleared out. This keeps the agent’s working context clean rather than cluttered with outdated context.

Merges and resolves conflicts. If an agent has accumulated duplicate entries or contradictory notes across different sessions, dreaming reconciles them into a coherent memory state.

Spots patterns. The agent looks across up to 100 prior sessions to identify recurring mistakes or inefficiencies in how it approaches tasks. It can then update its own preference files to avoid repeating those errors.

The practical result is an agent that gets more useful the longer it runs. According to multiple reports, Anthropic describes memory and dreaming together as forming “a robust memory system for self-improving agents.”

The feature is launching as a research preview that requires developer access. It is not automatically available to all Managed Agents customers yet.

Other Updates from Code with Claude 2026

Dreaming was not the only thing Anthropic announced. The company also released several additions to the Managed Agents developer platform:

Outcomes. A way for developers to define what a “successful” agent run looks like, giving the system a target to evaluate its own performance against.

Multiagent orchestration. Improvements to how multiple Claude agents coordinate with each other on complex workflows, making it easier to build systems where different agents handle different parts of a task.

Webhooks. Developers can now set up event-driven triggers so that agent activity can connect to other systems and workflows without requiring constant polling.

Together these updates suggest Anthropic is pushing the Managed Agents platform toward production-grade enterprise deployments rather than just proof-of-concept integrations.

What This Means for Business

The dreaming feature addresses one of the most common complaints about AI agents in production: they do not get better.

A human employee who makes the same mistake twice, gets feedback, and learns from it is more valuable on day 60 than on day one. Most AI agents today do not work that way. Each session starts fresh, with no accumulated understanding of what worked and what did not.

Dreaming is Anthropic’s attempt to close that gap. If an agent is handling customer intake, processing internal reports, or managing a recurring operational workflow, the ability to consolidate memory and self-correct between sessions has real value.

There are caveats worth noting. The feature is still in research preview, which means Anthropic is not making production-readiness claims yet. How much the self-improvement actually compounds in practice, and over what timeframe, remains to be seen.

But the direction is clear: Anthropic is building toward AI agents that accumulate organizational knowledge over time rather than resetting with every task. For businesses running agents on repetitive, knowledge-intensive work, that is a capability worth watching closely.

The Broader Picture

Managed Agents launched in public beta in April. At that point, Notion, Rakuten, Asana, and Sentry were among the early enterprise customers already running production deployments.

The addition of dreaming, outcomes, and orchestration improvements suggests Anthropic is moving quickly to close the gap between “agent that can complete a task” and “agent that improves at completing tasks.” That distinction will matter a lot for any business thinking about deploying AI across ongoing operational work.

If your business is evaluating AI agents for sustained, repeated workflows, the dreaming feature is worth tracking closely as it moves from research preview to general availability.

---

Enterprise DNA helps businesses deploy AI that creates lasting operational leverage. If you are thinking about AI agents for your business, book a discovery call with Sam McKay to talk through what is practical for your situation.