Anthropic Taught Claude to Dream — Tasks Got 6× Faster

What happens when an AI agent goes offline? For two years the honest answer was “nothing.” It forgets, you reload context, and you pay for the same lessons twice. On May 6, 2026, Anthropic shipped a feature that quietly breaks that pattern. It is called dreaming, and the first pilot customer saw task completion climb roughly six times.

What Dreaming Actually Does

Memory consolidation, scheduled

Dreaming is a scheduled background process that reviews an agent’s past sessions and its memory store, extracts patterns, and curates what survives into long-term memory. Anthropic describes it as a way to help agents “self-improve” between runs, the same way a junior analyst gets better not while they are typing but in the hours after they stop. Instead of throwing every transcript at the next session as raw context, the agent distills lessons — what failed, what worked, which tools paid off — and rewrites its own working memory accordingly.

Why this is a quiet architectural shift

Until now, “memory” in production agent systems mostly meant vector search over chat logs. Dreaming is different in two ways. First, it runs offline on a schedule rather than at request time, so it does not steal latency from the user-facing loop. Second, the model itself participates in curating memory — deciding what to keep, summarize, or discard — instead of an external retrieval layer guessing. That moves a core piece of agent intelligence from infrastructure code into the model itself.

Outcomes Replace Prompts as the Contract

A rubric the agent grades itself against

Outcomes is the second new feature, and it is probably the most undersold. Instead of writing a long prompt and praying, you describe what success looks like as a rubric — measurable criteria the agent uses to grade its own work before returning it. The agent runs, self-evaluates against the rubric, and iterates until the outcome is met or a budget is exhausted. Wisedocs, which uses Claude for medical document review, reported that adopting outcomes cut review time by 50%.

Webhooks finally make long-running agents useful

Tied to outcomes, Anthropic shipped webhook support: define an outcome, kick the agent off, and your service receives a callback when it is finished. This is the missing piece that turns a Managed Agent from a chat replacement into a backend worker — something you can hand a job to and walk away from, like a queue job, but with reasoning attached.

Multiagent Orchestration Without the Glue Code

A lead agent that delegates to specialists

The third release is a built-in orchestration tool that lets a lead agent break a job into pieces and delegate each one to a specialist subagent with its own model, prompt, and tool set. The lead agent picks who to call, in what order, and how to combine their outputs. Before this, teams spent weeks writing custom routing logic on top of LangGraph, CrewAI, or homegrown frameworks just to get this pattern working reliably.

Mix Opus for planning, Haiku for execution

Because each subagent can run its own model, the economic story changes too. You can spend Opus tokens on the planning step, dispatch Haiku for routine tool calls, and only escalate to Sonnet when a specialist subtask actually needs it. The lead agent enforces the boundaries. This is the first version of the pattern that ships configurable in the managed runtime rather than as glue code your team has to maintain.

What This Means for Builders

The agent stack is collapsing into the platform

Memory, planning rubrics, orchestration, and async callbacks are precisely the pieces that independent agent frameworks have been competing on for the last eighteen months. Folding them into Managed Agents pulls value out of the framework layer and into the model provider. For teams shipping agents in production, the buy-versus-build calculus on orchestration logic just shifted toward buy.

Availability and what to test first

Dreaming is in research preview, while outcomes, multiagent orchestration, memory, and webhooks are in public beta as part of Managed Agents. The fastest way to feel the difference is to take a repetitive agent workflow you already run, swap its prompt for a rubric-based outcome, and let dreaming run on the resulting session log for a week. The Harvey and Wisedocs numbers suggest the gains compound rather than appear in a single run.

Related

• Why Enterprises Are Quietly Leaving GPT for Claude
• He Raised $700M Before Showing Anyone the Product
• 40% of the Fortune 50 Are Already Using This
• Forget ChatGPT — SAP Just Bet $1B on Tables
• AI Just Found 10,000 Bugs Nobody Can Patch Yet

Sources

1. Anthropic — New in Claude Managed Agents: dreaming, outcomes, and multiagent orchestration
2. SD Times — New in Claude Managed Agents: dreaming, outcomes, and multiagent orchestration
3. Digital Trends — Anthropic just taught Claude to dream between tasks
4. 9to5Mac — Anthropic updates Claude Managed Agents with three new features