Anthropic Dreaming: Claude Agents Now Learn From Their Mistakes

Anthropic just gave its AI agents something no major lab has offered before: the ability to learn from their own history without any human intervention. The company announced a feature called "dreaming" on May 6 at its Code with Claude developer conference in San Francisco — and early results from customers suggest the productivity gains are real.

Legal AI company Harvey reported a roughly 6x improvement in task completion rates after implementing the feature. Medical document review firm Wisedocs cut its review time in half. These aren't benchmarks — they're production outcomes from teams that have been running Claude agents at scale.

What Dreaming Actually Does

Here's the problem dreaming is solving. Every AI agent runs with a limited context window — basically a cap on how much information it can hold in working memory at once. For a single conversation, that's manageable. But when agents work on long-running projects over days or weeks, they keep starting fresh. They repeat mistakes. They rediscover workflows they've already figured out. They lose institutional memory.

Dreaming is Anthropic's fix. It's a scheduled background process that reviews an agent's past sessions and memory stores, extracts patterns from them, and curates what's worth keeping. Think of it as the AI equivalent of a team debrief — except it happens automatically, across multiple agents, and at whatever cadence you set.

What makes it different from existing "compaction" techniques (where models summarize long conversations) is the cross-agent scope. Dreaming can pull patterns from an entire fleet of agents working on the same project. If multiple sub-agents keep making the same mistake on a particular type of task, dreaming surfaces that. If agents working independently converge on the same efficient workflow, dreaming preserves it for the whole team.

Who Controls the Process

You get two modes. Automatic mode lets the agent update its own memory files when it dreams, hands-off. Review mode requires a human to approve each batch of memory changes before they're applied.

You also control the trigger. Dreaming can run on a schedule — say, after every workday — or you can invoke it manually with the /dream command. For teams that want predictability over their agents' behavioral evolution, the review mode is the obvious choice. For teams running at scale who want continuous improvement without the overhead, automatic mode is the path.

Two More Features Hit Public Beta

Dreaming is currently in research preview, meaning developers need to request access. But Anthropic also used the May 6 announcement to push two previously preview-only features into public beta — available now to all Managed Agents users.

The first is outcomes. This lets you show an agent exactly what "good" looks like by providing a worked example of an ideal output. A separate grader agent then evaluates every output against that example. Anthropic's own tests show that using outcomes improves task success rates by up to 10 percentage points compared to standard prompting alone. For subjective work — writing in a brand voice, producing a specific kind of analysis — the gains can be even larger.

The second is multi-agent orchestration, which lets a lead agent break complex tasks into smaller jobs and delegate them to sub-agents. The Claude Console now shows exactly what each sub-agent did, step by step, so teams can audit agent behavior and catch problems early.

Why This Matters for Enterprise AI

The pattern Anthropic is building toward is significant. Claude Managed Agents already handle long-running workflows in legal, medical, financial, and engineering domains. Dreaming adds a compounding improvement loop on top of that: the more work an agent does, the better it gets at that specific type of work.

That's a meaningful moat for enterprises that have run Claude agents long enough to accumulate rich session history. A law firm that has run thousands of contract review sessions will have agents that have learned from all of them. A competitor starting fresh won't.

For Anthropic, this is also a way to deepen lock-in without building explicit switching costs. The value isn't just in the model — it's in the accumulated memory of how your team works.

The Broader Context

This release is part of a sprint of Managed Agents improvements Anthropic has been pushing since early 2026. In March, the company made multi-agent orchestration available in limited preview. In April, outcomes entered preview. Now both have graduated to public beta alongside the new dreaming capability.

The timing aligns with Anthropic's aggressive push into enterprise verticals — financial services, legal, healthcare — where agents that get smarter over time are worth far more than commodity AI completions.

In a final detail from the announcement: Anthropic also doubled usage limits for Pro and Max subscribers, from five hours to ten hours per session. For long-running agentic workflows, that matters.

What's Next

Dreaming is available in research preview today — developers can request access through the Anthropic developer console. Outcomes and multi-agent orchestration are in public beta and available to all Managed Agents customers without a waitlist.

Anthropic has said publicly that dreaming will graduate to full availability once the research preview phase is complete, though no timeline was given. For teams running production agent workflows right now, the outcomes and orchestration features are worth immediate evaluation — the task success improvements documented in Anthropic's tests are large enough to affect real business metrics.

The bottom line: AI agents that improve themselves between sessions are no longer a research concept. They're in production, and the early results are compelling enough to take seriously.