On Tuesday, Anthropic held its second annual Code with Claude developer conference in San Francisco. The headline announcement was a feature called "dreaming."
The name sounds abstract. The function is not.
Dreaming is a scheduled process that reviews an AI agent's past work sessions, identifies patterns - recurring mistakes, workflows that worked, preferences shared across multiple agents - and updates the agent's behavior going forward. Think of it as a nightly review meeting that the AI runs on itself, without you having to prompt it.
According to Anthropic, legal AI company Harvey saw task completion rates increase roughly 6x after implementing dreaming. Medical document review company Wisedocs cut its document review time in half. Netflix is using a related feature to process logs from hundreds of builds simultaneously.
Those are enterprise numbers. But the underlying system is available to any developer building on the Claude platform - and increasingly, that means tools you may already be using.
What "Dreaming" Actually Does (Plain English Version)
Every time you use an AI agent - a tool that takes instructions and acts on them autonomously - that agent runs a session. It might draft emails, pull data from a spreadsheet, research a question, or handle customer messages.
Normally, each session starts fresh. The agent does not remember what it did well or poorly last time.
Dreaming changes that. It sits above the session level and watches patterns across many sessions. It is not modifying the underlying model - it is not retraining the AI from scratch. Instead, it is curating memories and surface-level lessons that the agent can draw on the next time.
Anthropic's Alex Albert described it this way: when a person works through a complex task, they often zigzag before finding the most efficient path. By the end, they know the route. Dreaming records that route, so the next session starts closer to the destination.
Two Other Features That Matter for Business Users
Alongside dreaming, Anthropic moved two previously experimental features into public beta:
Outcomes - Lets you define what "success" looks like for an agent task, so the agent can course-correct if it drifts. Wisedocs used this to cut their document review time by 50%.
Multi-agent orchestration - Lets multiple Claude agents work on different parts of a task simultaneously instead of in sequence. If you have a research-then-write-then-format workflow, this can compress the timeline significantly.
Both are now broadly available to developers - meaning the tools built on top of Claude should start incorporating them over the coming months.
The Business Implication
Here is the honest version of what this means if you run a small business.
Right now, most small business owners using AI are still in the "prompt and hope" phase. They write a prompt, get a response, revise, repeat. The AI is useful but inconsistent. It does not get better at your specific business over time.
The direction Anthropic is heading - and where this release points - is toward AI that actually adapts to your context. An agent that has run 200 scheduling tasks for your business would, with dreaming, start to notice that you always prefer Tuesday mornings for external calls and adjust accordingly. Without you re-explaining it every time.
That kind of contextual memory is what turns AI from a fancy search engine into something closer to an actual assistant.
It is not here yet in consumer tools. But the infrastructure for it is being built now.
What to Watch For
If you use any AI platform built on Claude - that list includes Notion AI, Xero, certain Zendesk features, and a growing number of vertical SaaS tools - watch for announcements about "persistent memory" or "agent learning" over the next 6-12 months. That is the dreaming capability working its way into the products you already pay for.
If you are evaluating AI tools right now, ask vendors a simple question: does the AI get better the more I use it? An honest answer will tell you a lot about how they are building.
CEO Dario Amodei disclosed at the conference that Anthropic saw what he described as 80x annualized growth in revenue and usage in Q1 2026, far beyond its internal projection of 10x growth. API volume on the Claude platform is up nearly 70x year over year. The company that makes the AI your tools are probably running on is growing faster than almost anyone planned for.
That matters because the rate of new feature releases is likely to accelerate. Keeping up is genuinely worth your time.
Danny Kowalski covers AI tools and software for small businesses at The Useful Daily. Source: VentureBeat - Anthropic introduces "dreaming" (May 8, 2026); Anthropic Claude Managed Agents documentation.