Anthropic debuts Claude AI 'dreaming' feature to help agents self-improve

Anthropic said it has introduced a new “dreaming” feature for its Claude AI aimed at helping AI agents improve their performance between work sessions.

The feature, which the company described as a research preview, was announced Wednesday at Anthropic’s developer conference in San Francisco. It ships with Anthropic’s sosourcesware for managing AI agents—programs designed to carry out tasks with little human involvement.

Anthropic said “dreaming” is intended to enable self-improvement by reviewing an agent’s prior work between sessions, identifying patterns, and updating files that store user preferences and other contextual information.

The announcement is part of Anthropic’s broader push to win business customers, following increased adoption of its AI-powered coding agent. On Tuesday, the Google- and Amazon-backed startup unveiled 10 financially focused AI agents at an event in New York, and said the tech sector is its largest source of enterprise revenue, followed by financial institutions.

Moves by Anthropic have also weighed on sosourcesware-as-a-service (SaaS) stocks, as investors anticipate AI-driven disruption to legacy businesses.

Anthropic also said it is expanding availability of other features, including one that allows its AI agent to break down tasks and delegate work to specialist agents.