Dreaming in Claude Managed Agents is basically Anthropic’s way of giving your AI agents a good night’s sleep and a chance to learn from yesterday’s mess. Instead of starting every session with a clean slate, agents can now quietly review what happened, spot patterns, and update what they “know” so they get better over time without you hand-holding every step.

At its core, dreaming is a scheduled background process that goes over your agents’ past sessions and their memory stores, then distills what actually matters. Think of it like an editor going through a reporter’s notes after a long day: it pulls out recurring issues, useful tricks, and shared preferences, and throws away the noise. You can choose how much control you keep – let dreaming automatically update memory, or require human review before changes land.

What makes dreaming interesting is that it sees patterns a single agent, stuck in one context window, would miss. It can notice that across dozens or hundreds of runs, your agents keep hitting the same edge case, converging on the same workaround, or adapting to a team’s specific style and workflows. For long-running projects or setups where multiple agents are collaborating, that kind of “across sessions” perspective is exactly what’s been missing in most agent frameworks.

Dreaming also solves one of the big pain points with agent memory: bloat. As you let agents remember more, their memory can turn into a junk drawer where everything goes in and nothing comes out. Dreaming periodically reorganizes that drawer, curating and restructuring memory so what remains is high signal and actually useful. That keeps retrieval sharp, reduces confusion, and helps the system stay aligned with how your organization actually works today, not six months ago.

The feature is designed to pair tightly with Managed Agents’ built‑in memory system. Memory lets each agent capture what it learns while it works in real time: user preferences, domain quirks, successful workflows, failure cases that got fixed. Dreaming then kicks in between sessions to refine that raw material, pull out shared learnings across different agents, and keep the whole memory stack current and coherent. Together, they form a feedback loop where agents don’t just react; they gradually improve.

On the developer side, Anthropic is rolling dreaming out as a research preview inside Claude Managed Agents on the Claude Platform, with access gated through a request form. That “research preview” label matters: it signals that Anthropic wants teams to experiment, stress‑test the feature, and help shape how automatic self-improvement should behave in real products. It also means you should expect the APIs and best practices to evolve as they see what people build.

This self-improvement angle is already showing up in how early adopters are using Managed Agents. Legal AI company Harvey, for example, uses Managed Agents to coordinate complex drafting and document work, and dreaming helps those agents remember practical workarounds and tool-specific patterns between sessions, which translated into much higher completion rates in their tests. The key idea: instead of each task being a one-off, the system gradually accumulates institutional memory about what actually works in that environment.

From a product perspective, dreaming is Anthropic leaning into a very human pattern: teams get better when they stop and reflect. Most agent systems today are stuck in “always on, never reflecting” mode. You can prompt them better, give them tools, maybe chain them together, but they don’t really have a built-in rhythm of doing, then reviewing and improving. Dreaming formalizes that reflection cycle at the platform level, so you don’t have to reinvent it in every app.

For practitioners building serious workflows – things like log analysis, document review pipelines, or multi-step content production – this can be a big deal. Instead of relying purely on prompt engineering and manual monitoring, you get a structured way for your agents to internalize lessons over time. When patterns of failure show up, dreaming can surface them and help adjust memory so the same mistakes become less likely in future runs. When certain workflows consistently outperform others, those paths can effectively become the new default.

There’s also a governance angle here. Because you can decide whether dreaming’s memory updates apply automatically or go through review, teams can match the feature to their risk tolerance. Highly regulated environments might insist on human approval before anything changes persistent memory, while faster-moving teams might opt for more automation and treat the system’s behavior as something to watch in metrics and logs instead of gate manually. That flexibility will matter a lot as companies push agents deeper into real business processes.

Zooming out, dreaming is a signal of where agent platforms are headed: away from static “stateless API calls plus some tools” and toward living systems that accumulate experience and adapt. It sits at the intersection of memory, evaluation, and orchestration, and tries to encode a simple idea: your agents should actually learn from their own history. For teams investing in AI agents as long-term infrastructure, features like this are going to separate toy demos from systems that quietly get better month after month.