What’s in this article
- What Claude’s “Dreaming” feature is — how your agents now train themselves overnight.
- Why this matters for agency work — the shift from one-off builds to training appreciating assets.
- Here’s how I’d actually use this — a four-step playbook for safely activating and training a dreaming agent.
- What this changes for designer-run agency work — how to scope, price, and sell self-improving agents as a service.
- My $0.02: How I’d roll this out — a three-day plan to deploy your first self-training agent for a client.
🚀 Plug this into Claude Code or Claude Desktop
This spec contains a complete configuration for a Claude Managed Agent with the ‘Dreaming’ feature enabled. It includes a starter prompt, memory review guidelines, and a sample ‘golden log’ to kickstart the training process.
Want to turn this into a billable service? The Talk-to-Build community is where we break down how to productize these workflows. Or, book a working session and we’ll set up your first agent together.
The biggest friction in using AI agents has always been the training. You build an agent, give it a task, it makes a mistake, you correct it. An hour later, you give it the same task, and it makes the same mistake. The agent never learns. Until now.
Anthropic just shipped a feature for Claude Managed Agents called “Dreaming.” It’s a background process that lets your agents learn from your interactions while you’re offline. It reviews your chat history, spots the patterns in your corrections, and builds its own memory. It gets smarter while you sleep.
This isn’t a small tweak. The legal AI company Harvey tested this and saw task completion rates jump by 6x. For anyone building AI workflows for themselves or for clients, the game just changed. Your agents are no longer static tools; they’re junior team members that train themselves.
What actually shipped
“Dreaming” is a new setting on Claude Managed Agents. When you turn it on, a background process kicks in after you finish a session. This process scans the full transcript of your conversation. It looks for moments where you corrected the agent’s output or rephrased a command to get a better result.
From these corrections, it creates what Anthropic calls “memories.” These are simple, internal rules the agent will follow in the future. For example, if you constantly edit an agent’s emails to be less formal, it might create a memory like: “When writing emails, use a more casual tone and avoid corporate jargon.”
You get full control over this process. You can set it to “Automatic,” where it learns and applies memories on its own. Or you can use “Review-Only” mode, where it proposes new memories every morning for you to approve or reject. This lets you guide its learning without giving up control, which is how I’d recommend everyone start.
DAY (you work) NIGHT (Claude dreams) MORNING (return)
────────────────── ────────────────────── ──────────────────
You chat with Claude Reviews transcript New task, same prompt
You correct mistakes Spots correction patterns Claude applies what
You close the laptop Creates internal memories it learned overnight
↓ ↓ ↓
one session passive learning compounding speed
This feedback loop turns every interaction into a training session. You’re not just getting work done; you’re actively making the agent better for tomorrow without any extra effort.
Your agent is now a junior team member that trains itself overnight, for free.
Why this matters for agency work
For designers and agency owners, this solves the single biggest problem with selling AI automation: brittleness. Clients don’t want a tool that breaks if they phrase a request slightly differently. They want a reliable system. Dreaming makes agents less brittle and more reliable over time.
This fundamentally changes the value of an AI agent. It’s no longer just a piece of software you build once. It’s an asset that appreciates. The longer a client uses it, the smarter it gets, and the more valuable it becomes to their business. It learns their specific jargon, their preferred formats, and their unique workflows.
This creates a powerful moat. Once an agent has six months of learning baked in, a client is not going to switch to a competitor’s off-the-shelf tool and start the training process from scratch. The agent you built becomes a core piece of their operational infrastructure. That’s a much stickier, higher-value relationship than just building a simple automation.
Here’s how I’d actually use this
This is a powerful feature, but you need to roll it out carefully. Turning a self-modifying AI loose on client work without guardrails is a bad idea. Here is the exact four-step process I’d use to deploy a dreaming agent safely.
- Activate Dreaming in ‘Review-Only’ Mode. When setting up a new Managed Agent, I’d immediately turn on Dreaming, but I’d select the option that requires manual approval for all new memories. This gives me all the learning benefits without the risk of the agent learning a bad habit automatically.
- Feed it a ‘Golden Log’ of Perfect Sessions. The agent learns from the conversations it has. To speed things up, I’d start by feeding it a perfect example. I’d have one long, clean session where I walk it through a task correctly from start to finish. For example, I’d use the logs from my team onboarding new members to the Talk-to-Build community to train a new community manager agent. This gives it a strong foundation to learn from.
- Make Memory Review a Daily 10-Minute Habit. Each morning, I’d check the agent’s dashboard for proposed memories. The UI shows you the proposed rule and the conversation that triggered it. I’d spend a few minutes approving the good ones and rejecting the bad ones. This is the most important part of the process—it’s like giving feedback to a new employee.
- Switch to Automatic After a Week of Clean Reviews. Once I’ve gone a full week where I’ve approved 95% or more of the agent’s proposed memories, I’d feel comfortable switching it to Automatic mode. By then, I’ve guided its initial learning, and I can trust it to continue improving on its own. I’d still check in weekly, but the intense daily oversight is no longer needed.
This process turns you from a simple user into a trainer. You’re not just prompting; you’re shaping the agent’s long-term behavior. *If you can talk it, you can build it.*
What this changes for designer-run agency work
This isn’t just a new feature to mention to clients. It changes how you should structure your entire AI service offering. Here are three immediate shifts I’m making in how I scope this work at MK-Way.
Your core offer shifts from ‘building agents’ to ‘training agent teams.’ You’re no longer selling a one-time setup. You’re selling a process of continuous improvement. The initial build is just step one. The real value is in the ongoing training and refinement, which you are uniquely positioned to manage for the client.
This justifies a higher-value, performance-based retainer. A static automation might be worth a $300/month maintenance fee. An agent that gets measurably better every month is worth much more. I’d structure this as a $500-$1,500/month “Agent Performance Management” retainer. The goal is to tie the fee to the agent’s improving capabilities, like its speed or accuracy on key tasks.
It dramatically lowers the barrier to deploying complex agents. Previously, building an agent that could handle nuance and edge cases required weeks of complex prompt engineering. Now, you can start with a simple v1 agent and let it learn the complexity over time through real-world use. This means you can take on more sophisticated projects and deliver value to clients faster.
The conversation with a client is no longer about what an agent can do on Day 1. It’s about what it will be able to do on Day 90, after it has learned their business inside and out.
My $0.02 — How I’d roll this out for a design business
If you want to start selling this service today, here’s a simple three-day plan. This is how you go from reading this article to having a new, billable service.
Day 1 — Isolate one repetitive, high-friction workflow in your own business. Don’t start with a client. Start with yourself. For me, a great example is processing new member applications for the Talk-to-Build community. It’s a structured task that involves checking payments, sending welcome emails, and adding users to the right groups. I’d document the exact steps I take manually.
Day 2 — Build a v1 agent and run the workflow for a day. I’d create a new Claude Managed Agent with a simple prompt that covers the basics of the task. I’d turn on Dreaming in ‘Review-Only’ mode. Then, I’d spend the day using that agent to process real applications, correcting its mistakes as I go. By the end of the day, I’ll have a rich log of real-world interactions for it to learn from.
Day 3 — Review the memories, create a case study, and pitch your first client. In the morning, I’d review all the memories the agent proposed overnight. I’d approve the good ones, and now my agent is officially smarter. I’d take screenshots of the process and write a one-page summary: “How I built an agent that automates member onboarding and gets smarter every day.” That one-pager is now the sales asset I can send to my best client.
This is the exact playbook I use to develop new services. I build it for myself first, live the workflow, productize the process, and then teach it. It’s the fastest way to create offers that are grounded in real, practical experience.
FAQ
What does this cost?
Dreaming is a feature of Claude Managed Agents, which is part of the standard Claude API pricing. You pay for the compute used during the overnight processing, but it’s designed to be very efficient. Anthropic says it’s a small fraction of the cost of the interactive sessions themselves.
Is my data used to train Anthropic’s main models?
No. The memories created by Dreaming are specific to your agent and are not used to train Anthropic’s foundation models. Your conversational data remains private to your account.
What’s the difference between Dreaming and fine-tuning?
Fine-tuning is a heavy, expensive process where you retrain the entire model on a large dataset. Dreaming is a lightweight, continuous process that adds a ‘memory’ layer on top of the existing model. It’s faster, cheaper, and works in near-real-time.
Can the agent learn bad habits?
Yes, which is why starting in ‘Review-Only’ mode is critical. If you accidentally teach it the wrong thing, you can go into its memory and manually delete the incorrect rule.
How many conversations does it need to learn?
It can start learning from the very first conversation. The more you interact with it and correct it, the faster it builds a useful set of memories.
What kind of tasks is this best for?
It’s ideal for repetitive tasks that have a lot of small, specific rules. Things like customer support triage, data entry from unstructured documents, content formatting, and scheduling are all great use cases.
Want help applying this?
Four ways to go deeper:
- Build with Builders. Join the Talk-to-Build community to learn how to Earn money with AI, Download our AI Skills, Advance your business, and learn to build real assets — AI-native websites, cinematic AI video, agent-driven workflows — that you can sell to SMBs who want the outcomes but don’t have time to learn the skills.
- 1-on-1 working session. Skip the friction. Book a screen-share with me — bring a real problem, leave with a working piece of it.
- Done-for-you. MK-Way builds AEO-ready websites, apps, and AI agent workflows for design agencies and founders who want it shipped fast.
- Quick question. DM me on Instagram or connect on LinkedIn. I read every message.
This post is part of the AI Pulse atomic series. If you commented “AGENTS” on one of my videos — this is the breakdown. Sources: Anthropic Blog.
Last updated: 2026-06-01.