On May 6, Anthropic shipped a feature called Dreams for its managed agents platform. The feature lets agents learn between sessions.

Dreams lets AI agents review their own past sessions between tasks. They analyze what happened. They identify recurring mistakes. They extract patterns from workflows that worked. They restructure their memory so the next session starts from a better baseline than the last one.

This is an agent that reflects on its work, learns from what went wrong, and shows up better prepared tomorrow. Anthropic’s own description: “Dreaming surfaces patterns that a single agent can’t see on its own, including recurring mistakes, workflows that agents converge on, and preferences shared across a team.”

Read that last part again. Preferences shared across a team. The agent learns from its own history and from what every agent on your team has experienced.

Why This Matters More Than the Headlines Suggest

Most coverage of Dreams focused on the technical architecture. Scheduled review sessions. Memory curation. Pattern extraction. All accurate. All missing the point.

The point is this: Anthropic just built the feedback loop that separates a tool from a collaborator.

A hammer behaves the same way each time you pick it up, regardless of what you built yesterday.

A collaborator needs memory. A collaborator needs to know what worked last time, what failed, what you prefer, and what the team has learned collectively. Without that feedback loop, every interaction starts from zero. You explain the same context. You correct the same mistakes. You watch the same errors repeat.

That is exactly what most enterprise AI looks like right now. Every session is a fresh start. Every agent is amnesiac. The humans carry the institutional knowledge. The AI carries none.

Dreams starts to change that architecture, though its memory remains imperfect and incomplete.

What a Workflow Looks Like When Agents Actually Learn

Here is the difference in practice.

Without Dreams: You deploy an agent to handle IT service desk tickets. It resolves tickets based on its training. When it gets something wrong, a human corrects it. The next ticket starts with no memory of that correction. The agent makes the same class of error again. Your team builds workarounds. They write documentation the agent never reads. The cycle repeats.

With Dreams: The agent reviews its past sessions. It notices it misrouted network issues to the wrong team three times last week. It notes that the resolution pattern for VPN tickets changed after the infrastructure update on Tuesday. It restructures its memory to reflect these patterns. The next shift starts from that updated baseline. The human corrections from last week are now part of the agent’s working knowledge.

That resembles onboarding. The same process you use to bring a new team member up to speed. Except the agent does it to itself, between shifts, without anyone scheduling a training session.

Anthropic shipped two companion features alongside Dreams that make this concrete. Outcomes lets you set up evaluator agents that grade outputs against ideal examples. Early results show a 10 percentage point improvement in success rates compared to standard prompts. Multi-agent orchestration lets a lead agent decompose complex tasks and distribute them to sub-agents, with full visibility into who did what.

Combined, these three features describe something recognizable. A team. With roles. With feedback. With a shared memory of what works.

The Standard Is Moving

Here is what business leaders should be watching.

ServiceNow just deployed autonomous AI specialists that handle entire business processes. AWS made its MCP server generally available, giving agents secure access to cloud infrastructure. Google renamed its entire cloud platform around agents last week. And now Anthropic is shipping the memory and learning layer that lets those agents build on prior work.

These announcements describe infrastructure layers assembling underneath the same shift. Some teams are already running agent squads that get measurably better every week. Everyone else is still explaining context to stateless chatbots every Monday morning.

The gap between those two groups is operational. And Dreams is the kind of feature that makes it compound. An agent that learns can improve with each day it runs. By week eight, the difference between a learning agent and a stateless one becomes a difference in accumulated capability.

Dreams is in research preview. It requires an access request. It is early. None of that changes the trajectory. Within 18 months, every major agent platform will have some version of this. The only variable is which teams will have 18 months of accumulated agent learning when that happens, and which will be starting from scratch.

The collaborator model requires you to answer that question.