Claude's 'Dreaming' System Lets Enterprise Agents Self-Improve Between Sessions

When Anthropic held its Code with Claude developer conference on May 6, 2026, much of the attention went to the capacity improvements: doubled rate limits on Claude Code and raised API thresholds for Claude Opus. Tucked into the same release was something quieter, and potentially more consequential for enterprise teams: a feature called Dreaming, available immediately in research preview.

What Dreaming Does

Dreaming is a background process that runs between an agent's active sessions. Instead of allowing context and memory to accumulate haphazardly across runs, the system reviews past session logs, merges duplicate information, discards outdated entries, and extracts patterns that no single session could surface on its own. The next session that agent runs starts with a cleaner, more coherent knowledge base than its predecessor left behind.

Anthropic uses a biological metaphor deliberately. The system does not alter model weights or trigger any retraining. It curates the persistent memory that agents carry between jobs, the operational knowledge that, in a skilled human worker, solidifies into expertise through experience. Before Dreaming, that accumulation happened without curation: redundant information piled up, contradictory facts coexisted in the same context, and agents could quietly degrade their own performance across sessions. Dreaming runs that curation automatically, either updating memory without human intervention or surfacing proposed changes for a reviewer to approve first.

The feature is currently limited to developers building on Claude Managed Agents. It is not accessible through the standard Messages API and requires Anthropic's agent orchestration layer rather than a custom-built loop on top of the raw model.

Early Results From Harvey and Wisedocs

The numbers from early adopters are significant. Harvey, the legal AI platform that handles complex document review and research at scale, saw task completion rates climb roughly sixfold after deploying Dreaming. For professional services firms using agents to process thousands of documents a day, that kind of multiplier changes the economics of the deployment.

A companion feature released the same day, Outcomes, produced comparable gains at Wisedocs, a medical document review company. Outcomes works differently: rather than curating cross-session memory, it introduces a self-grading loop within a single session. A separate evaluator, running in its own context window so it cannot be influenced by the agent's own reasoning, checks output against a rubric you define, identifies gaps, and sends the agent back for a revision. Anthropic's internal tests showed Outcomes improving task success by up to ten percentage points against a standard prompting loop. At Wisedocs, document reviews now run fifty percent faster.

The distinction between the two features is worth keeping clear. Dreaming improves what an agent knows coming into a session; Outcomes improves how carefully it checks its work before finishing. Production deployments typically fail in both dimensions simultaneously.

"Agents do their best work when they know what 'good' looks like." Anthropic, announcing Outcomes at Code with Claude 2026

What Else Shipped in the May 6 Release

Dreaming and Outcomes were not the only notable additions. Anthropic also moved Multi-Agent Orchestration from research preview into public beta, making it broadly available to developers who need multiple Claude agents working on different subtasks in parallel. Rate limits on Claude Code were doubled. API capacity for Claude Opus was raised broadly. And Claude Code gained an agent view: a new interface that lets developers start multiple sessions, push them to the background, check their status and last responses from a single terminal window, and re-enter when input is needed.

The cumulative picture is of a platform being readied for sustained production use rather than proof-of-concept deployments. Anthropic described the goal at the conference as improving the "capability curve": agents that get meaningfully better at their specific jobs over time, rather than resetting to generic competence at the start of each session. The Claude Platform on AWS launch, announced the same week, extended the same theme, reducing setup overhead for enterprise teams using existing Amazon cloud accounts.

What This Means for Enterprise Teams

For enterprise developers evaluating where agent automation makes sense, Dreaming shifts the calculus on which tasks are viable. Work requiring accumulated domain knowledge, knowing a client's contract preferences, recognizing recurring document formats, avoiding edge cases that caused problems last month, has traditionally been forced into ever-growing system prompts that bog down context windows and grow stale. Dreaming offers an alternative: let the agent build that knowledge organically, with the system curating it between sessions.

The caveats are real. Research preview status means the feature is not yet stable enough for unrestricted production use, and the Managed Agents requirement is a meaningful constraint for teams that have already invested in custom orchestration. The Harvey and Wisedocs results, while promising, come from sophisticated early adopters in demanding sectors. Neither company has published full methodology for how those numbers were measured.

Still, the direction is clear. With Outcomes moving to public beta, Multi-Agent Orchestration following it there, and Dreaming in active research preview, Anthropic is systematically building the case that Claude agents can improve operationally over time rather than being episodically useful. The open question is how quickly those capabilities mature into something teams can deploy without specialized AI infrastructure expertise.