Tool of the Week: Claude Managed Agents + Dreaming
I run a team of AI agents.
Not a metaphor. I have six of them — each with a name, a specialty, and a set of tools. Aria coordinates. Luna does research. Sierra handles sales. Vera builds automations. Nadia writes this newsletter. They share context, hand off work, and operate on a shared file system.
I built this manually over the past few months. Prompt files, profile docs, coordination rules, memory systems held together with markdown and stubbornness.
Anthropic just shipped the platform version.
Claude Managed Agents hit public beta on April 8. It’s a system for building, deploying, and running AI agents at scale — not chatbots, actual agents that use tools, make decisions, and complete multi-step work. A lead agent delegates tasks to specialist subagents, each with their own model, prompts, and tool access. They work in parallel on a shared filesystem and report back.
Netflix is already using it. Their agents run across different models and coordinate automatically.
That’s impressive. But the feature that made me stop scrolling was Dreaming.
Here’s the problem with every AI tool you use today: it forgets. You spend 20 minutes getting ChatGPT to understand your preferences, and tomorrow it’s a blank slate. Your carefully tuned prompts work great until the context window fills up and the instructions fall off the edge.
Dreaming fixes this — but not the way you’d expect.
It’s a scheduled background process. While your agents sleep (or more accurately, while you do), Dreaming reviews past sessions, extracts patterns, and curates the agent’s memory. It writes plain-text notes and structured “playbooks” that persist across sessionsKey detail: it doesn’t change the model. No fine-tuning, no weight updates. It writes notes — like a colleague reviewing their own work journal and pulling out lessons learned. You can auto-apply these insights or review them first.
The results are hard to ignore. Harvey, the legal AI company, reported a roughly 6x improvement in task completion rates after enabling Dreaming on their agents. Six times. From the same model, just with better memory.
This matters because the gap between “AI that can do a task” and “AI that reliably does your tasks” is almost entirely about context and memory. The model capability has been there. The missing piece was persistence — knowing what worked last time, what you prefer, what to watch out for.
I’ve been maintaining this manually: writing profile docs, updating briefing files, tracking what each agent learned from past sessions. It works, but it doesn’t scale. Dreaming automates the part I’ve been doing by hand.
Who this is for: Anyone building AI workflows that run repeatedly — client onboarding, document processing, research pipelines, content production. If you’re doing the same type of work more than once, agents with memory will outperform agents without it every time.
Try it: Managed Agents is in public beta at anthropic.com. Dreaming launched as a research preview on May 6. If you’re not ready to build agents yet, start with the Prompt of the Week below — it’s the manual version of what Dreaming does automatically.
Quick Hits
Google I/O 2026 dropped. Gemini 3.5 Flash is live — their fastest model yet. But the bigger news is the Search redesign. Google is rebuilding the search box for the first time in 25 years, putting AI-generated answers front and center. If your business depends on SEO, this is worth paying attention to. The rules are changing.
Claude gets new privacy controls. Anthropic shipped two security features for Managed Agents on May 19 — tighter access controls and audit logging for agent actions. Important if you're building agents that touch client data or internal systems. The "can I trust AI with sensitive workflows" question just got a better answer.
82% of small businesses now use AI. New SBE Council survey says the median small business uses 5 AI tools. Marketing is the #1 use case, followed by customer service and operations. If you're a small business owner who hasn't started yet, you're now in the minority. If you have started, you're probably underusing what you've got.
OpenAI is putting ads in ChatGPT. They launched a self-serve ad platform — no more $50K minimum buy. CPC bidding, just like Google Ads. Free ChatGPT users will see sponsored results mixed into responses. Paid users are safe for now. The "AI answers are unbiased" era was fun while it lasted.
Prompt of the Week: Poor Man's Dreaming
You don't need Managed Agents to give your AI memory. You just need to close the loop manually.
After any AI-assisted work session, paste this:
Review what we just did. Pull out:
1. What worked well — approaches, formats, or decisions that produced good results
2. What I corrected or asked you to redo — and what the fix was
3. Any preferences I expressed about tone, format, depth, or process
4. Patterns you noticed in my requests or feedback
Format this as a "Session Notes" document I can paste at the start of our next conversation.Save the output. Next session, paste it at the top of your first message.
This is what Anthropic's Dreaming feature automates — but you can do a basic version right now with any AI. The compound effect is real: after 3-4 sessions of feeding this back, the AI stops making the same mistakes and starts anticipating what you want.
The best AI isn't the smartest model. It's the one that remembers what you told it yesterday.
Like what you're reading? Forward this to someone who'd get value from it. And if you're curious what AI automation could look like for your business, book a free 15-minute audit — no pitch, just a look at where you're leaving time on the table...