Earlier this month at its developer event, Anthropic announced "dreaming" — a feature for Claude Managed Agents that reviews an agent's past sessions during idle time, extracts patterns from what went well and what didn't, and updates the agent's memory accordingly. The internet immediately decided this was either the dawn of conscious AI or marketing nonsense. Both takes miss what's actually useful about it.
What it actually is
Strip the metaphor and dreaming is a scheduled background process. Between active work, the agent looks back over its previous sessions and memory, flags recurring mistakes and team preferences, merges duplicate notes, and removes outdated ones. The model weights — the actual "brain" of the AI — are not changing. What changes is the working memory the agent reads before its next task. It's closer to an employee writing themselves a better playbook on Friday afternoon than to anything magical.
Why this matters for business agents
The single biggest reason real-world AI agents underperform is that they make the same mistakes repeatedly. They forget last week's feedback. They re-discover the same answer for the third time. Dreaming directly attacks that pattern by consolidating what the agent has already learned. Anthropic cited the legal AI company Harvey seeing task completion rates jump roughly sixfold after enabling it — that result deserves scrutiny, but it does point at the right kind of problem.
What's still your job
You decide whether the agent updates its memory on its own or whether you review the changes first. For anything touching client work, regulated decisions, or pricing, you want the review gate on. The whole point of memory is that the next task inherits it — which means a bad memory propagates the same way a good one does, and a quiet update to an agent's playbook is exactly the kind of change that can drift unnoticed for weeks.
The honest caveat
Dreaming is in research preview, not general availability. Real ROI numbers will take months to show up in customers' own data, not curated case studies. Treat this as a meaningful capability worth piloting if you already run agent workflows — not as a reason to start running them.
Your takeaway
The interesting question this raises for any SMB running AI agents is the right one: how does your agent actually learn from its mistakes today? If the answer is "it doesn't," features like dreaming are about to make a real difference between vendors. Ask about it on your next demo.