Anthropic introduces 'dreaming' feature to enable Claude AI agent self-improvement

Anthropic has introduced a novel "dreaming" feature for its Claude AI, aiming to bolster the capabilities of its autonomous agents for business customers.

The San Francisco-based startup, which enjoys significant backing from Google and Amazon, unveiled the tool on Wednesday during its annual developer conference.

This "dreaming" capability allows AI agents to engage in a self-improvement loop by reviewing past sessions and identifying patterns that may have been overlooked during active tasks.

Available as a research preview within the company’s software for managing agents, the feature functions by reviewing an agent's work in between sessions.

It unearths recurring mistakes and updates long-term memory files that store user preferences and essential context.

This development follows a surge in popularity for Anthropic’s AI-powered coding assistant, as the firm pivots toward providing scalable infrastructure for large-scale enterprise operations.

The announcement coincided with a New York event where Anthropic showcased 10 new financially focused AI agents designed for tasks such as auditing and credit memo drafting.

CEO Dario Amodei noted that the company’s growth rate has significantly exceeded expectations, necessitating a recent deal with SpaceX to utilise the Colossus 1 data centre for increased computing power.

As Anthropic expands the availability of features that allow agents to delegate tasks to specialists, the market has seen a notable impact on legacy software-as-a-service (SaaS) stocks.