Anthropic adds a “Dreaming” feature for its AI agents

In short: Anthropic announced a new “Dreaming” feature that lets its AI agents review what they just did and try to do better next time.

What happened

Anthropic introduced “Dreaming” at its developer conference in San Francisco. The feature is part of the company’s tools for building and running AI agents.

An AI agent is software that can carry out a task in several steps, like opening websites, reading files, and reporting back. You can think of it like a helper that follows a checklist on your computer.

Anthropic says “Dreaming” works by looking back at an agent’s activity log, which is basically a transcript of what the agent did. It then tries to pull out patterns and useful lessons that could improve future performance. In Anthropic’s description, “memory” is what the agent saves while it is working, and “dreaming” is what it does between sessions to refine that saved information.

Why it matters

The announcement also highlights a broader debate about how AI features are named. Many companies use human-sounding terms like “reasoning,” “thinking,” and “memory,” even though these systems do not think or feel the way people do.

Researchers have warned that giving machines human-like labels can change how people judge them, including how much they trust them and who they blame when something goes wrong. For everyday users, the key point is simple: these tools can be helpful, but the names can make them sound more human, and more capable, than they really are.

Source: Wired