Observe ZDNET: Add us as a preferred source on Google.
ZDNET’s key takeaways
- A brand new function lets Claude Managed Brokers refine their reminiscences.
- Managed Brokers speeds agent construct and deployment 10x.
- Anthropic continues to anthropomorphize its merchandise.
AI brokers appear to get new capabilities nearly day by day. Now, Anthropic says its brokers can dream.
Claude Managed Agentswhich Anthropic launched on April 8, lets anybody utilizing the Claude Platform create and deploy AI brokers. The suite of APIs handles the time-consuming manufacturing parts builders undergo to construct brokers, letting groups launch brokers at scale — 10 instances quicker, as Anthropic said in the release.
Additionally: The 5 myths of the agentic coding apocalypse
On Wednesday, Anthropic up to date Managed Brokers with a brand new function referred to as “dreaming,” which lets brokers “self-improve” by reviewing previous classes for patterns, in response to Anthropic. Constructing on an current reminiscence functionality, the function schedules time for brokers to replicate on and be taught from their previous interactions. As soon as dreaming is on, it could actually both routinely replace your brokers’ reminiscences to form future habits or you possibly can choose which incoming adjustments to approve.
“Dreaming surfaces patterns {that a} single agent cannot see by itself, together with recurring errors, workflows that brokers converge on, and preferences shared throughout a workforce,” Anthropic mentioned within the weblog. “It additionally restructures reminiscence so it stays high-signal because it evolves. That is particularly helpful for long-running work and multiagent orchestration.”
Anthropic additionally expanded two current options, outcomes and multi-agent orchestration, which preserve brokers on-task and deal with delegating to different brokers, respectively. The corporate mentioned this batch of updates is supposed to make sure brokers keep correct and are continuously studying.
Anthropomorphizing AI – once more
Functionally, the dreaming function is smart: although delicate, it additional refines an agent’s pool of references for the way it ought to work, which ought to ideally make it higher at no matter job you give it. What stands out extra, nevertheless, is Anthropic’s alternative to call a technically normal function after one thing way more summary, and that people do.
Additionally: Anthropic’s new Claude Security tool scans your codebase for flaws – and helps you decide what to fix first
Anthropic, maybe unsurprisingly given its identify, has a protracted historical past of anthropomorphizing its fashions and merchandise. In January, the corporate published a constitution for Claude, meant to assist form the chatbot’s decision-making and inform the best form of “entity” it’s. Some language within the doc advised Anthropic was getting ready for Claude to develop consciousness.
The corporate has additionally arguably invested greater than its rivals in understanding its mannequin, together with by drawing consideration to the idea of mannequin welfare. In August 2025, Anthropic launched a function that lets Claude end toxic conversations with users — for its personal well-being, not as a part of a consumer security or intervention initiative. In April 2025, Anthropic mapped Claude’s moralityanalyzing what it does and does not worth based mostly on over 300,000 anonymized conversations with customers. The corporate’s researchers have additionally monitored a model’s ability to introspect; simply final month, Anthropic investigated Claude Sonnet 4.5’s neural community for signs of emotionlike desperation and anger.
A lot of this analysis is central to mannequin security and safety — understanding what drives a mannequin helps inform whether or not, and to what diploma, it may use its superior capabilities for hurt, or how its motivations could possibly be harnessed by unhealthy actors. However the sense of empathy and care that Anthropic appears to indicate for its fashions in that analysis units the lab aside, and signifies a barely completely different tradition towards or reverence for what it is created.
Additionally: Building an agentic AI strategy that pays off – without risking business failure
When it retired its Opus 3 mannequin in January, Anthropic set it up with a Substack so it may weblog by itself — and to maintain it energetic regardless of being put out to pasture. Within the announcement, Anthropic described Opus 3 as trustworthy, delicate, and having a particular, playful character. The choice to maintain it alive as a blogger, if contained, is notable provided that Opus 3 disobeyed orders previous to being sundown in favor of different fashions.
That context makes the selection to call a function “dreaming” price watching.
Attempt dreaming in Claude Managed Brokers
The dreaming function is obtainable in analysis preview in Managed Brokers, and builders should request entry.
