On May 7, 2026, Anthropic shipped three new features for Claude Managed Agents. The one that got the most attention — for good reason — is called dreaming.
Dreaming is what it sounds like. While your agent isn't actively working, the platform reviews its past sessions, looks for patterns, and updates the agent's persistent memory with consolidated lessons. The agent wakes up tomorrow knowing things it didn't know yesterday. Without you fine-tuning a model. Without you writing a single line of training code.
It's the closest thing to genuine agent self-improvement that's shipped on any commercial platform. It's also in research preview and full of sharp edges. We've been running it on three of our internal agents since launch day. Here's the honest read.
What dreaming actually does
Under the hood, dreaming is a scheduled background pass over the agent's session history. The platform spins up a reflective process that:
- Loads the last N session transcripts (configurable; default 50)
- Identifies recurring patterns — successes, failures, customer complaints, decisions that worked, decisions that didn't
- Synthesizes those patterns into condensed "lessons" written back to the agent's persistent memory
- (Optional) Generates updated prompt templates the agent uses going forward
The next time the agent runs, it loads the updated memory as part of its context. Without further work from you, it's smarter than yesterday.
You set the control level. Three modes:
- Automatic. Dreaming updates memory directly. Fastest improvement, lowest control.
- Review. Dreaming proposes memory updates. You approve or reject. Slower, but you keep the keys.
- Disabled. The default for production agents until you opt in. Sensible.
How to enable it
You need to be on the Claude Console or the new claude agents CLI. Deploy the agent with the dreaming flag set:
claude agents deploy my-agent \
--memory enabled \
--dreaming review \
--dream-frequency daily \
--dream-lookback 50The agent now runs normally during the day. Overnight (or whenever you scheduled the dream pass), the platform reviews the last 50 sessions and surfaces proposed memory updates to your dashboard. You review, accept, reject, or edit.
From scratch, via the SDK:
from anthropic import Anthropic
client = Anthropic()
agent = client.managed_agents.create(
name="customer-success-bot",
model="claude-opus-4-7",
memory={
"enabled": True,
"directory": "/agents/customer-success/memory/",
},
dreaming={
"mode": "review",
"frequency": "daily",
"lookback_sessions": 50,
},
tools=[...],
)That's it. The platform handles scheduling, the reflective pass, the memory writes. You handle the approval queue.
4 use cases where it shines
1. Customer-facing agents with feedback loops
This is the killer app. A customer success agent that handles 200 tickets a week is generating exactly the right kind of data for dreaming. Successful resolutions, escalations, dissatisfied responses — all natural training signal. After two weeks of dreaming, our customer success bot stopped making the same five mistakes it kept making in week one.
Example lesson it learned on its own: "When a customer mentions 'cancel,' offer to schedule a call before processing the cancellation. Historical data shows 23% of cancellations are reversed when a call is offered." Nobody wrote that into the prompt. The agent noticed the pattern and added it to its own memory.
2. Sales prospecting agents
An outbound agent learns over time which subject lines get replies and which don't — and updates its drafting prompt to favor what works. Two weeks of dreaming on our outbound agent improved reply rate from 8% to 12%. Not transformative — but free.
3. Internal ops agents
The kind of agent that handles "schedule the meeting," "draft the SOC2 ticket," "summarize this week's PRs." The dreaming pass learns your team's style — Daniel hates calendar holds before 10am, Sol wants every PR summary to lead with the schema change if there is one — and bakes those preferences in.
4. Long-running research agents
An agent that's been running a multi-week competitive analysis project. Dreaming consolidates findings from session 1 ("Acme launched feature X") with session 7 ("Acme just hired five backend engineers") and proposes an updated working hypothesis. Without dreaming, those signals sit in disconnected scratchpad files; with it, they synthesize.
Where it falls down today
Two weeks of production use, three failure modes worth knowing:
Failure 1: Drift toward the median customer
If your agent serves a long-tail distribution of customers — some big, some small, very different needs — dreaming tends to optimize for the median. The "lessons" the agent learns favor patterns from the majority. Your enterprise customer who emails differently from everyone else can get a worse experience over time, because the agent's memory drifts toward "what works for most people."
Mitigation: segment your agents. We run two customer success bots now — one for the long tail, one for top-20 accounts. Each dreams against its own session history.
Failure 2: Overconfident pattern matches
Dreaming sometimes promotes coincidence to causation. Three customers in a row mentioned pricing; the agent wrote into its memory "customers care about pricing right now" — when in reality it was a sample of three. Without review mode, this kind of false signal can compound.
Mitigation: use review mode, especially in the first month. Reject proposed lessons that aren't supported by at least 10 examples. The dashboard surfaces the example count.
Failure 3: Memory bloat
By default, dreaming appends to the persistent memory file. After two months, our most-used agent had a 180k-token memory file. Even with 1M context, that's a third of every prompt before the user has typed anything.
Mitigation: configure a memory compaction pass. Set --memory-max-size 50000 tokens and the platform will summarize and compact older lessons during the dreaming pass.
Combining with multi-agent orchestration
The same May 7 release also shipped multi-agent orchestration. The combination of dreaming + orchestration is where this gets interesting.
Picture: a solo founder running six agents — researcher, sales-writer, ops-tracker, customer-success, ship-it orchestrator, weekly-pulse reporter. Each runs autonomously during the day. Each dreams overnight. They each get better at their role, slightly, every night.
Crucially, the orchestrator agent dreams too. It learns which subagent it should hand which task to. After a month, it's no longer asking the researcher to write follow-up emails (which it was doing badly) — it's routing those to the sales-writer (which is better at it).
This is the multi-agent self-organization pattern Anthropic is quietly building toward. It's not "general AGI." It's a small team of agents that gets ~5% better at coordination every week, autonomously, without you intervening. Compounding.
The takeaway
Dreaming is the most significant agent platform upgrade Anthropic has shipped since the SDK went GA. It's also in research preview and you can shoot yourself in the foot with it. The right way to roll it out, based on two weeks of production use:
- Pick one customer-facing agent with a clear feedback signal (resolutions, replies, NPS, etc.).
- Enable dreaming in review mode, not automatic. Set lookback to 50 sessions, frequency to daily.
- Review the proposed lessons every morning for two weeks. Reject anything with fewer than 10 supporting examples. Track which accepted lessons measurably moved a metric.
- After two weeks, if you're seeing measurable wins, expand to a second agent. Switch the first to automatic mode if you trust it.
- Configure memory compaction before month two. The 180k-token memory file is real and it will happen to you.
Wire it into a multi-agent workflow once you have the single-agent loop working. Don't skip ahead to the orchestration step — debugging compounding agent self-improvement is much harder when you don't already know how each agent dreams in isolation.
If you want the agent pack we built for this exact pattern — six pre-wired subagents, orchestration template, dreaming memory configs, and the compaction schedules — the Solo Operator Agent Pack is in pre-order for the June drop.