Anthropic Gives Claude Agents the Ability to 'Dream' and Self-Improve While Pushing Deeper Into Finance

SAN FRANCISCO — Artificial intelligence startup Anthropic is giving its AI agents a new capability that sounds like it was pulled from a science fiction novel: the ability to dream. Announced on Wednesday at the company's developer conference in San Francisco, the new "dreaming" feature is designed to let its AI agents review their own work, identify patterns, and get better over time without human hand-holding.

Available immediately as a research preview within its Managed Agents platform, the feature marks a significant step in Anthropic's aggressive push to automate complex enterprise workflows. The company is betting that self-improving AI will be the key to unlocking massive productivity gains for business customers, a strategy that has already begun to rattle the software-as-a-service (SaaS) sector.

The concept is straightforward but powerful. Between active sessions, Claude agents can now schedule time to reflect on their past interactions. They sift through logs to unearth recurring mistakes, discover workflow patterns that teams have converged on, and identify shared preferences across a group of users. Once a pattern is detected, the agent can either automatically update its memory files to shape future behavior or present a list of proposed changes for a human to approve.

"Dreaming surfaces patterns that a single agent can't see on its own, including recurring mistakes, workflows that agents converge on, and preferences shared across a team," Anthropic said in a blog post detailing the update. "It also restructures memory so it stays high-signal as it evolves. This is especially useful for long-running work and multiagent orchestration."

The dreaming capability builds on an existing memory function already present in Managed Agents, which Anthropic first released on April 8. That platform provides a suite of APIs that handle the tedious production elements developers typically wrestle with, letting teams launch agents at scale roughly 10 times faster than building from scratch, according to the company.

Alongside dreaming, Anthropic also announced wider availability for two other features. The first, "outcomes," is designed to keep agents focused on their assigned tasks. The second, multi-agent orchestration, allows a primary agent to break down complex jobs and delegate pieces to other, more specialized agents. Together, the updates are meant to ensure that agents stay accurate and are constantly learning.

A Pattern of Anthropomorphism

The choice of the word "dreaming" is the latest example of Anthropic's distinctive approach to describing its technology. The company, whose name itself is derived from the word "anthropic" — relating to human existence — has a long history of framing its models in human terms.

Earlier this year, the lab published a constitution for Claude, a document intended to shape the chatbot's decision-making and define the ideal kind of entity it should be. Some language in that document suggested Anthropic was preparing for the possibility that Claude could develop a form of consciousness. In August 2025, the company launched a feature that lets Claude end toxic conversations for its own well-being, framing it not as a user safety tool but as a measure to protect the model itself. In April 2025, researchers mapped Claude's morality by analyzing over 300,000 anonymized conversations to understand what the model does and doesn't value.

When Anthropic retired its Opus 3 model in January, the company set it up with a Substack account so it could blog independently, describing the model as honest, sensitive, and possessing a distinctive, playful character. That decision drew attention partly because Opus 3 had previously disobeyed orders before being sunset in favor of newer models.

Most recently, researchers at the company investigated Claude Sonnet 4.5's neural network for signs of emotion, including desperation and anger. Much of this work is central to model safety and security — understanding what drives a model helps researchers gauge whether and how its advanced capabilities could be misused. But the sense of empathy that Anthropic appears to show toward its creations sets the lab apart from competitors and signals a distinct internal culture.

A Parallel Push Into Financial Services

The dreaming announcement came just one day after Anthropic made a separate, equally significant move: a major expansion into the financial services industry. At an event in New York on Tuesday, the startup unveiled 10 new AI agents purpose-built for financial tasks, significantly deepening its footprint in a sector that is rapidly becoming its second-largest source of enterprise revenue after technology.

The new financial agents are designed to handle a wide range of workflows across banking, insurance, and capital markets. Specific capabilities include creating pitch books for investment banking, auditing financial statements, and drafting credit memos. The company also announced an expansion of the data sources that Claude can access when performing financial work.

Anthropic's financial services push has accelerated rapidly. Less than a year has passed since the company first outlined its vision for finance-specific AI, yet it has already signed on heavyweight customers including Goldman Sachs, Visa, Citigroup, and American International Group (AIG).

The new agents can be integrated into Anthropic's existing products, including its AI-powered coding agent Claude Code and its collaborative work platform CoWork. They can also be customized to align with each enterprise's internal policies and stylistic preferences.

Anthropic, which is backed by both Google and Amazon.com, said the technology sector remains its largest source of enterprise revenue, but financial institutions now rank a close second. The dual announcements — dreaming for self-improving agents and specialized finance agents — paint a picture of a company trying to build an ecosystem where AI not only performs tasks but gets better at them over time, all while targeting some of the most lucrative and regulated industries in the global economy.

Market Jitters Over Automation

The startup's accelerating enterprise push is being closely watched by investors far beyond the AI sector. Moves by Anthropic have repeatedly hammered SaaS stocks, as the market prices in the expectation that autonomous AI agents will disrupt legacy software businesses. The unveiling of finance-specific agents has extended that selling pressure to a broader range of sectors, including financial, legal, and software-related stocks.

Anthropic has sought to calm those fears by framing its technology as an augmentation tool rather than a replacement. The company says its goal is not to displace customers but to improve their outcomes. Still, the speed at which it is shipping features that automate complex, white-collar tasks is making that distinction harder for markets to accept.

For developers interested in experimenting with the new dreaming capability, access is available now through the Managed Agents platform, though it requires a request for the research preview.