Anthropic’s efforts to build self-improving AI agents have led to a new anthropomorphism technique called “Dreaming.”
The renowned AI Lab announced the new process at its developer conference on Wednesday. “Dreaming” is part of Anthropic’s commitment to overhauling software engineering and other knowledge work with tools that increase autonomy. This new functionality will be included in the initial Claude Managed Agents product.
This technique aims to improve the system’s memory by performing evaluations between sessions. It reviews old behaviors and looks for patterns, helping agents establish better ways of working and reduce mistakes.
For now, the “dreaming” process has begun as a research preview, and developers must apply for access.
Anthropic’s revenue has soared in recent months as software engineers embrace the company’s Claude Code service and related services, which help developers deploy agents to drive long-term coding projects. The startup is looking to expand these capabilities beyond software engineering into fields such as finance and law.
By having agents remember and learn from their previous work, Anthropic agents can become more accurate and productive over time, increasing their value to paying customers.
The new release also highlights a splashy essay published Monday by Jack Clark, Anthropic’s co-founder and head of policy. Clark claimed in the Import AI newsletter that there is a 60% chance that Frontier AI models will be able to autonomously train successor models by the end of 2028.
He cited the current wave of research and new product developments from Frontier Labs. Clark didn’t mention “dreaming,” which falls under the company’s efforts to turn its AI models into a workhorse of partially self-managing agents.
“The significant increase in the length of time that AI systems can operate independently correlates neatly with the explosion of agent coding tools,” Clark said in his essay. “This is the commercialization of AI systems that can work on behalf of humans and operate independently over long periods of time.”
Also Wednesday, Anthropic moved two managed agent tools from research preview to public beta. One uses results in a rubric format to guide agents. Second, work can be delegated to multiple subagents.
Any tips? To contact this reporter via email, please specify the following address: scouncil@insider.com or Signal, Telegram, or WhatsApp (415-757-8198). Use a personal email address, non-work WiFi network, and non-work device. Here’s a guide to sharing your information securely.