Anthropic introduces “dreaming,” a system that allows AI agents to learn from their own mistakes

Anthropic announced a series of updates to its Claude agent platform, including a new capability called “dreaming” that allows AI agents to learn from their past work sessions and improve over time, a step toward self-correcting and self-improving AI systems that companies have been demanding before letting agents handle production tasks, Venture Beat writes.

At the same time, the company moved from the experimental stage to the testing stage of two other features: “outcomes” (outcomes) and multi-agent orchestration, which will be available on the Claude platform. All three capabilities attempt to solve what Anthropic says are the most difficult problems with AI agents: keeping them accurate, helping them learn, and preventing them from becoming obstacles in complex, multi-step tasks.

Promising results

Companies that have used them so far have reported significant results. Law firm Harvey reported that the completion rate of tasks increased sixfold using “dreaming”, and Wisedocs, a company specializing in the evaluation of medical records, announced that the time required was reduced by 50% by using outcomes.

Anthropic's announcement comes at a time of accelerated ascent for the company. According to the company's head, Dario Amodei, in the first quarter of 2026, Anthropic saw an 80-fold increase in revenue and usage of its services, well above the 10-fold increase that the company's management expected.

AI agents “take notes”

“Dreaming” is a programmed process that evaluates the AI ​​agent's past work sessions, extracts patterns and action from them, and manages its memory so that agents improve over time. This reveals things that the agent cannot see on its own: repetitive mistakes, activities that multiple agents perform independently, and shared preferences within a team of AI agents.

Alex Albert, who coordinates research at Anthropic, said that by “dreaming” AI agents are “taking notes” that they can use in the future and that can be observed and evaluated by human observers. All the “memories” thus stored can be inspected from the outside, and smarter models have improved the quality of these notes. “They learn to write better notes for what they will be in the future,” Albert said.

Alienation without human involvement

An Anthropic team demonstrated how the three new features work using a fictional spacecraft, “Lumara,” which is supposed to launch ground-based drones to the surface of the Moon for mining. The team set up a multi-agent system with three specialists – a commander agent responsible for the overall success of the mission, a detector agent who identified the most suitable landing sites, and a navigator agent who handled the safe flight and landing of the drones.

Initially, selenization was simulated at six hypothetical sites, with encouraging but imperfect results. Afterwards, a “dreaming” session was started from the Claude Developer console. Overnight, the dream agent evaluated past similar sessions and wrote a detailed manual for aselination. When the simulations were repeated the next day, this time based on the manual, the results improved significantly.

A problem that will need to be fixed

The demonstration showed how the three functions work simultaneously. “Multi-agent orchestration” broke down complex tasks into smaller tasks to specialized agents in independent context windows, and the results were better than when entrusting all these tasks to a single AI model.

“Effects” provided the rubric against which another agent could evaluate the results of each work session, and “dreaming” drew lessons from past experiences to improve future performance. Virtually the process of improving AI agents is carried out without human intervention.

Albert admitted that making an AI model verify itself is problematic. One of the issues is that repeated self-checking and self-correction over a large number of sessions leads to degradation of results over time, but Anthropic is working on fixing this as well.

But the ability of AI agents to consult their own experiences creates the conditions for artificial intelligence systems to be powerful enough for companies to delegate tasks of major importance to them, notes Venture Beat.