Anthropic Lets Claude Agents Dream to Learn From Their Own Mistakes

Harvey's engineers didn't change their prompts. They didn't swap models. They turned on dreaming for their Claude-powered legal agents and watched task completion rates jump roughly 6x.

That result is what Anthropic is leading with as it rolls out three new capabilities for Claude Managed Agents: dreaming, outcomes, and multiagent orchestration. The company announced all three on May 6, 2026, two of them newly available to all developers, and one entering a closely watched research preview.

The timing matters. AI agents are becoming production infrastructure at serious companies, not just prototypes on developer laptops. That shift is creating a hard requirement: agents that improve over time without constant human retraining. Dreaming is Anthropic's answer to that need, but the announcement was really three coordinated answers at once.

How Dreaming, Outcomes, and Orchestration Work

Dreaming is a scheduled background process that reviews an agent's past sessions and memory stores, extracts patterns, and restructures memory to keep signal quality high. It surfaces insights that no single agent session could see on its own: recurring mistakes, workflows that multiple agents converge on independently, and preferences shared across a team of agents.

The mechanism works differently from standard memory. Regular session memory captures what happened in a given conversation or task run. Dreaming works at a higher level of abstraction. After a set of agent sessions run, the dreaming process wakes up, reviews what happened across all of them, identifies what was consistent or consistently wrong, and distills that into refined memory the agent carries into future sessions. Think of it like the consolidation that happens during human sleep: raw experience gets compressed into durable knowledge.

Two control modes exist. Developers can let dreaming update memory automatically, useful for high-trust production workflows, or they can require manual review before any memory update takes effect. The second option makes the feature safer to deploy in regulated environments where any change to agent behavior needs explicit sign-off. Dreaming is available in research preview. Anthropic says the company may introduce breaking changes during preview with at least one week's notice and recommends against using it for critical or sensitive workflows during this phase.

The outcomes feature solves a separate problem: how to measure whether an agent's output is actually good, without trusting the same reasoning process that produced it. Anthropic's approach is to let developers write a rubric describing what success looks like, then run a separate grader against output in its own context window. The grader evaluates the result without seeing the agent's reasoning chain, so its assessment isn't anchored to how the agent justified its own answer.

When output falls short, the grader doesn't just flag it as wrong. It identifies specifically what needs to change and the agent takes another pass with that feedback. Anthropic reports +8.4% improvement on document generation and +10.1% on presentations in testing. Outcomes moved from experimental to public beta with this announcement.

The third feature, multiagent orchestration, also moved to public beta. A lead agent takes a complex task, breaks it into parts, and delegates each part to a specialist subagent. Each specialist has its own model, prompt, and tool configuration. They run in parallel on a shared filesystem and contribute findings back to the lead agent's broader context. Every step is traceable in the Claude Console. The combination of parallel execution and full visibility is what makes the Netflix log-analysis use case possible: one orchestrator divides hundreds of builds across specialist subagents, synthesizes the patterns, and surfaces only what's worth acting on.

These three features aren't separate product updates. They're components of a shared architecture. Outcomes gives you measurement. Dreaming takes those quality signals and aggregates patterns across many runs, encoding what was learned into refined agent memory. Multiagent orchestration expands the surface where dreaming can operate, producing richer session data faster.

The Harvey, Netflix, and Wisedocs Numbers

Harvey uses Claude Managed Agents to coordinate complex legal work: long-form drafting, document creation, and multi-step research tasks. These are workflows where small errors compound quickly and where output quality depends heavily on understanding how a specific legal team works, its preferred citation styles, draft structures, and client-specific terminology rules. Completion rates increased roughly 6x in testing after Harvey enabled dreaming.

The result makes sense in context. Legal work is full of recurring patterns. When an agent recognizes those patterns across hundreds of sessions and improves its baseline understanding of them, it performs better without requiring prompt engineers to manually codify every rule. Dreaming does that recognition automatically, at a scale no manual process can match.

Netflix's engineering team is testing a different use case. Their platform team built an analysis agent that processes logs from hundreds of builds across different sources. Multiagent orchestration lets the agent run batch analysis in parallel, and dreaming helps surface only the patterns worth investigating rather than requiring an engineer to read everything. The scale is different from Harvey's, but the core value is similar: the agent learns what matters across many sessions, not just within one.

Wisedocs, a document processing company, reports that their document reviews run 50% faster while maintaining team standards alignment. Faster is a useful headline, but the standards alignment piece is the harder one to get right, and dreaming is what Anthropic says is producing that alignment over time.

These three cases span legal AI, developer platform analytics, and document processing. They're not a homogeneous dataset, but they share one characteristic: all three involve complex, repetitive, slightly-variable work where human experts get noticeably better over months on the job. That's the pattern profile where dreaming is designed to add value.

Deployment Considerations and Open Questions

The clearest near-term implication is that early deployment data matters more than it used to. Before dreaming, session outcomes were useful feedback for prompt engineers, mostly as qualitative signals. With dreaming enabled, those outcomes are direct inputs to agent improvement. Teams that deploy early and log outcomes carefully will have better-performing agents sooner.

This changes the calculus on pilot scope. It used to make sense to run a small pilot with limited volume just to validate quality. With dreaming, a pilot with low session volume won't produce enough pattern data to trigger meaningful improvement cycles. Think about minimum volume thresholds and design pilots accordingly.

The outcomes feature changes review workflows. If grader-based iteration handles the first several revision cycles automatically, human reviewers can focus on whether the output actually serves the user, not on copy-editing or structural fixes. That shift requires teams to redefine what 'done' means for agent output, and to calibrate grader rubrics carefully before production use. A poorly written rubric causes the agent to iterate toward technically-correct-but-wrong output, so rubric design is now a first-class skill.

Multiagent orchestration adds deployment complexity and cost. Parallel subagents each consume context and model capacity. On budget-constrained deployments, running many subagents simultaneously can spike costs quickly. Build usage monitoring into orchestrator deployments from the start. And when an orchestrator delegates to a set of subagents, a subagent failure needs graceful handling so it doesn't silently corrupt the final output. Build explicit failure modes before scaling.

Research preview status means Anthropic can break the dreaming API with one week's notice. Teams building production workflows on top of dreaming need to isolate it from critical paths until it exits preview. Memory updates that run automatically could also drift in unexpected directions if session volume is too low or too noisy. A small sample of bad sessions could encode incorrect patterns. Automatic mode requires monitoring and alerting if agent behavior shifts unexpectedly.

For a current view of how these capabilities compare across major agent platforms, our Agent Tools Comparison resource covers the market with regular updates as platforms ship new features.

The dreaming research preview is the piece most worth watching. Once it exits preview, the question becomes whether it can scale beyond legal AI and document processing to broader categories of knowledge work. The cases Anthropic is highlighting share specific characteristics. Not every agent workflow will look like Harvey's.

For multiagent orchestration, the practical ceiling depends on how well developers can build specialist subagents that are genuinely narrow and reliable. Generalist agents pretending to be specialists don't produce the gains the orchestrator pattern promises.

This also connects to the broader shift toward agentic workflows that Andrej Karpathy described at AI Ascent 2026. In that framing, which we covered in our analysis of the shift from vibe prompts to agent workflows, the next competitive differentiator isn't prompt quality alone. It's the quality of systems that improve agent behavior at scale. Dreaming fits directly into that thesis.

The combination of all three features, especially when dreaming is stable enough for production use, creates something without a clear historical precedent in AI tooling: agents that improve automatically, grade their own work with independent evaluation, and scale through coordinated subagent teams. Whether that produces compounding quality gains in real enterprise deployments is the most important open question in the Managed Agents story right now.