Anthropic Dreaming: How Claude Managed Agents Learn Between Sessions

What Anthropic Actually Shipped

On May 6, 2026, Anthropic announced three capabilities for Claude Managed Agents at its Code with Claude event: dreaming, outcomes, and multi-agent orchestration. Of the three, dreaming is the most architecturally significant — it is the mechanism by which a Claude agent can observe its own behavior across sessions and automatically generate institutional knowledge for future runs.

Dreaming is not a metaphor. It is a scheduled background process that activates between agent sessions, reads the agent’s full activity log and existing memory store, surfaces recurring mistakes and successful patterns, and then writes new plain-text notes and structured “playbooks” into the memory layer. The underlying model weights are never touched. Instead, future sessions inherit richer context — effectively benefiting from everything prior sessions learned without any developer having to manually annotate or retrain.

According to Anthropic’s official announcement, users can configure dreaming to update memory automatically or to present proposed changes for human review before they are committed. This gives enterprise deployments the oversight they need while still allowing agents to compound their learning at software speed.

The timing matters: dreaming ships alongside outcomes and multi-agent orchestration, meaning developers can now combine a self-improving memory layer (dreaming) with structured quality grading (outcomes) and parallel task execution across specialist subagents. These three features together represent a qualitative shift from single-session tool-use to something closer to an organizational learning system.

Three Signals Hidden in the Architecture

1. Harvey’s 6x Completion Gain Reveals the Real Cost of Stateless Agents

Harvey, the legal AI company that processes documents and supports attorneys at large law firms, reported that task completion rates climbed roughly 6x after implementing dreaming. The explanation from Harvey’s team was specific: agents had been losing workarounds and learned preferences between sessions — for example, how to handle particular file type quirks or client-specific formatting standards. Each new session started cold, re-discovering the same friction points that previous sessions had already resolved.

According to Anthropic’s managed agents release notes, dreaming solved this by capturing these workarounds as playbook entries. The next session reads them before starting, treats the edge case as known territory, and moves forward without re-testing what already works. For Harvey’s legal workflow — where precision and throughput are both critical — the compounding effect of dozens of such micro-learnings produced the 6x figure. This signals that the greatest immediate value of dreaming is not in domains where agents regularly succeed, but in precision domains where edge cases cause silent failure or rework.

2. Wisedocs’ 50% Speed Gain Shows Outcomes and Dreaming Are Designed to Stack

Wisedocs, a medical document review company, cut review time by 50% using a different Anthropic feature announced the same day: outcomes. Outcomes allow developers to write rubrics defining what a high-quality output looks like. A separate grader model evaluates the agent’s work against those rubrics in its own context window, identifies gaps, and prompts the agent to self-correct before the output is ever delivered.

Anthropic’s internal benchmarks reported +8.4% success improvement on DOCX file processing and +10.1% on PPTX generation when outcomes were enabled. Wisedocs’ 50% efficiency gain came from eliminating the human review cycles that had previously caught errors the agent was producing at the output stage — errors that outcomes now catch programmatically before delivery. The significance for architects evaluating Claude Managed Agents: dreaming and outcomes are complementary, not redundant. Dreaming improves behavior over time. Outcomes enforces quality at every individual output. Stacking both is the path to enterprise-grade reliability.

3. Multi-Agent Orchestration Signals That Session-Level Dreaming Is a Stepping Stone

The third capability announced alongside dreaming is multi-agent orchestration: a lead agent that delegates tasks to specialist subagents, each with individual models, prompts, and toolsets, working in parallel on a shared filesystem. Netflix is already using this pattern to process logs from hundreds of builds simultaneously — a task that was previously sequential and bottlenecked.

The architectural implication is that dreaming’s memory layer was clearly designed with orchestrated systems in mind. When subagents accumulate session history and write playbooks, the lead agent can read those playbooks to understand each specialist’s strengths and constraints before delegation. This turns multi-agent coordination from a blind routing problem into an informed one. The practical horizon for enterprise CTOs is 12-24 months: current multi-agent systems still require careful task decomposition by a human architect, but dreaming is building the institutional memory that will eventually reduce that design burden.

What This Means for Enterprise AI Teams

1. Audit Your Agent Failure Modes Before Enabling Dreaming

Before activating dreaming on a production agent, enterprise teams should run structured analysis of where their agents currently fail or require human correction. The reason is architectural: dreaming surfaces patterns from existing session logs. If the existing logs reflect unreviewed errors (agents that made the same mistake repeatedly without correction), dreaming risks encoding those mistakes as playbook guidance.

The safest onboarding path is to run the agent in outcomes mode first — let the grader flag and fix errors until the output quality is stable — then enable dreaming to capture those corrected patterns. The sequence is: quality gate first, memory accumulation second. Reversing this order risks building a self-improving agent that self-improves in the wrong direction.

2. Map Your Memory Governance Policy Before You Scale

Dreaming writes new memory entries autonomously. In regulated industries — financial services, healthcare, legal — those memory entries constitute a form of operational procedure: they influence how the agent handles client data, compliance checks, and sensitive decision logic. Most organizations do not have a memory governance policy because most AI deployments do not generate self-authored procedures.

Teams should define, in advance: who can approve new memory entries before they are committed, how long memory entries persist, and what triggers a full memory audit. The Claude Console provides visibility into delegation and execution order across multi-agent systems, which is the right starting point for building that audit trail. Organizations that skip this governance step risk accumulating undocumented operational logic inside an agent’s memory store — the AI equivalent of undocumented code in a production system.

3. Benchmark Against Your Current Baseline Within 30 Days of Enabling Dreaming

The Harvey 6x figure and Wisedocs 50% gain are compelling, but they reflect specific workloads in specific domains. Neither number should be treated as a baseline expectation for a different deployment. The right comparison is each team’s own pre-dreaming baseline.

Establish measurable task completion rates, error rates, and time-per-task metrics before enabling dreaming. Re-measure at 30 days and 90 days. This is the only way to quantify the actual gain for a given workload — and the only data that will support a business case for scaling the deployment. According to Anthropic’s May 2026 announcement, dreaming is in research preview (not general availability), so early benchmarks also serve as product feedback that may influence how Anthropic develops the feature.

The Structural Lesson

The deeper significance of dreaming is not the Harvey or Wisedocs results — impressive as they are. It is what the feature architecture reveals about how Anthropic is conceptualizing the agent reliability problem.

Until now, the primary strategies for making AI agents more reliable were better prompts, more capable models, and human-in-the-loop oversight. Dreaming introduces a fourth strategy: structured self-observation. An agent that can review its own track record, identify what worked and what didn’t, and encode that knowledge for its successors is no longer purely a function — it becomes an institution. It starts to exhibit the same knowledge compounding that makes experienced human teams faster than new ones.

The implications are asymmetric across industries. In domains with high document volume and structured quality criteria — legal, medical, financial compliance — dreaming’s compounding effect is fastest because the failure patterns are dense and repetitive. In more creative or context-sensitive domains, the gains may be slower and harder to measure. For enterprise architects evaluating Claude Managed Agents in May 2026, the practical question is not whether dreaming is useful in principle, but whether the organization’s AI workloads are concentrated enough, and the failure patterns repetitive enough, to benefit from institutional memory accumulation within a 12-month horizon.

The feature is in research preview. The organizations that start building the governance infrastructure now — oversight workflows, memory auditing, outcome rubrics — will be the ones positioned to scale when it reaches general availability.

Frequently Asked Questions

Sources & Further Reading

• New in Claude Managed Agents: dreaming, outcomes, and multiagent orchestration — Anthropic
• Anthropic introduces “dreaming,” a system that lets AI agents learn from their own mistakes — VentureBeat
• Claude Managed Agents Dreaming Explained — Build Fast with AI
• Anthropic Launches Dreaming for Claude Agents at Code with Claude 2026 — Let’s Data Science
• Claude agents now dream between sessions — Medium