TOOLS & SKILLS
Claude Just Learned How to Dream 🛠️
Anthropic dropped three updates last week that all point at the same thing — agents that watch themselves, fix themselves, and improve over time. Here’s the.
What Just Happened
Anthropic released three new features for Managed Agents on the Claude Platform: Dreaming, Multiagent Orchestration, and Outcomes. They sound unrelated until you read them together. The whole package is one bet: agents that don’t just do what you tell them, but watch themselves, fix themselves, and improve on their own.
Most of this is shipped through the developer-facing Claude Platform, so you might not be using these features directly today. But the phrase “AI agents” is about to be everywhere, and the people who already understand the structure win the next 18 months.
Dreaming
Dreaming is a scheduled process that runs in the background. While the agent is idle, it goes back through its recent sessions and memory stores, finds patterns — recurring mistakes, workflows it converges on, preferences shared across a team — and curates that into plain-text notes and structured “playbooks” the agent can reference next time.
It is not retraining the model. Anthropic doesn’t touch model weights here. What changes is the agent’s playbook — the document the agent reads when it sits down to work. So the next morning, the agent is “smarter” because its operating notes got rewritten overnight.
Multiagent Orchestration
A lead agent breaks a big job into pieces and assigns each piece to a specialist agent — each with its own model, prompt, and tools. The specialists work in parallel on a shared filesystem and report back. The lead agent stitches everything together.
Practical example: instead of one agent trying to handle research + writing + design + QA in one context window, you have four specialists each handling their lane, all running at the same time. Faster, cleaner outputs, less context-window bloat.
Outcomes
You write a rubric describing what success looks like. The agent works toward it. A separate grader (in its own context window) evaluates the output against your criteria. If the output doesn’t meet the rubric, the agent keeps working — revising, retrying, refining — until it does.
In practical terms: instead of asking an agent to “write the email” and accepting whatever comes back, you tell it “write an email that hits these 6 criteria” and the agent self-corrects until the grader signs off.
What This Means If You’re Not a Developer
All three of these features ship through Anthropic’s developer surface (the Claude Platform / Console / API). You probably won’t flip them on personally this month. But the direction is clear:
• Memory is becoming structural. Agents that remember across sessions are the new default, not a hack.
• Multi-step work is becoming a team move. One-prompt-do-everything is going away.
• Self-correction is becoming the bar. “First draft and stop” output is going to feel embarrassing within 12 months.
Translation: the agents you’ll be using inside ChatGPT, Claude.ai, and every productivity app you touch in 2026 are about to get noticeably better. The people building with Claude now are the ones who’ll know how to operate them when they show up in mainstream tools next quarter.
5 Simple Agents You Can Build Today
You don’t need the developer features to start. Every one of these can be built inside Claude Code as a Project with a context file, a workflow prompt, and a memory directory. None require an engineering background.
Connects to your Gmail. Every morning, surfaces the 5 emails that actually need a response, drafts a reply for each, and ignores the rest. Memory: which senders matter most, which threads you’re tracking. Saves 30 min a day for most people.
Pulls numbers from one platform you care about (Shopify, Stripe, Klaviyo, GA4 — pick one). Builds the same weekly report every Monday morning so you’re never opening dashboards again.
Reads your calendar + your priorities doc + your last 7 days of work and tells you what to focus on today. Memory: what you committed to last week, what slipped, what’s urgent. The closest thing to a chief of staff you can build in a weekend.
Scrapes the accounts you respect, analyzes what’s working on the platforms you post on, and generates 5 specific content ideas tailored to your niche every week. Memory: what you’ve already posted, what flopped, what flew.
If you sell something: reads your reviews, support tickets, and comments. Surfaces themes — what people love, what they complain about, what language they use. The cheapest market research you’ll ever do.
The real unlock of AI agents is not full autonomy, but human-in-the-loop systems design. Structure your agent inputs with explicit constraints, negative prompts, and automated test checkpoints. When building tools, keep token consumption lean by caching system prompts and isolating tasks into specialized sub-agents.