Claude agents learn from experience and now work in parallel

Last week's Code with Claude conference in San Francisco produced three announcements UK businesses using Claude can act on straight away.

Anthropic revealed that Claude Managed Agents, which moved into public beta in April, now has three additional capabilities: Multiagent Orchestration, Outcomes, and Dreaming. The same event confirmed that all paid Claude plans now carry doubled rate limits, with peak-hour restrictions removed for Pro and Max subscribers.

What Multiagent Orchestration does Running a complex AI task in Claude has, until now, meant one agent working through steps in sequence. With Multiagent Orchestration, a lead agent breaks a large job into subtasks and hands each one to a specialist agent with its own model, prompt, and tools. Those specialists work in parallel on a shared filesystem and contribute their outputs back to the lead agent's overall picture.

For a UK business, this changes the economics of complex automation. A task that previously required a single Claude session to research, draft, and format something sequentially can now be split, with each element running at the same time. The ceiling on what a managed agent can realistically complete in one run is now substantially higher.

What Dreaming actually means Dreaming is the most unusual of the three features and probably the one with the longest tail of practical value.

At a set interval, a Dreaming-enabled agent reviews its past sessions, extracts patterns, and updates its own memory. It can surface recurring mistakes, identify workflows that consistently produce good results, and learn preferences shared across a team, all without you having to rewrite a prompt or rebrief the agent manually. You retain control: Dreaming can either update memory automatically or surface proposed changes for your review before anything is committed. Any business running recurring AI workflows, such as weekly reporting, monthly data pulls, or routine document processing, will find that the agent responsible for those tasks gradually becomes more accurate without additional oversight from the team managing it.

Outcomes adds a quality check The third feature, Outcomes, lets you write a rubric describing what good output looks like. A separate grader evaluates the agent's work against that rubric in its own context window, so it is not influenced by the reasoning that produced the original output. When something falls short, the grader identifies what needs to change and the agent tries again.

This is a practical quality control mechanism, particularly useful in contexts where output needs to meet a defined standard, such as compliance documents, client-facing reports, or structured data exports.

Rate limits: the immediate news Alongside the agent features, Anthropic confirmed that Pro, Max, Team, and Enterprise plans all now carry doubled rate limits, effective immediately. Peak-hour reductions have also been removed for Pro and Max users. The additional capacity comes from a new computing agreement with SpaceX, giving Anthropic access to over 220,000 GPUs at the Colossus 1 data centre in Memphis. For UK teams, the practical effect is more consistent throughput during business hours. Teams running Claude Code heavily between 09:00 and 17:00 should notice fewer capacity-related slowdowns.

How Adevious AI sees this At Adevious AI, we build client workflows on Claude and use Claude Code as a central part of how we work. The Multiagent Orchestration update shifts how we approach larger automation tasks. Where we previously designed a single long-running agent to handle research, drafting, and formatting in sequence, we can now split those responsibilities across parallel specialists. The output arrives faster and, because each agent is configured for a narrower job, tends to be more consistent. Dreaming is something we are planning to test across client deployments that run on a regular cycle. For any workflow that repeats weekly or monthly, the prospect of an agent that improves its own performance through repetition, rather than requiring prompt revision from our team, represents a genuine efficiency gain.