Story perspectives

Anthropic Introduces Dreaming, Improves Agent Success, Speed 50%

5/7/2026

Summary

Anthropic Dreaming

Anthropic announced Claude Managed Agents now have a “dreaming” feature that reviews past sessions.
Dreaming can run automatically or require manual approval before updating memory.
Anthropic added an outcomes loop that uses developer rubrics and a grader, raising task success by 10 points (8.4% for docx, 10.1% for pptx).
Multi-agent orchestration lets a lead agent assign specialist agents, with Netflix and Wisedocs reporting a 50% speed boost in document reviews.

Summary

Anthropic Revenue Surge

Anthropic's revenue surged as developers adopted Claude Code and related AI services.
The firm plans to launch Managed Agents in finance and legal, not software.
Dreaming remains a research preview that requires developers to apply and use an email on Wi-Fi.
Co-founder Jack Clark said there is a 60 % chance AI models will train successors autonomously by 2028.
Critics say the “dreaming” label anthropomorphizes AI, referencing Philip K. Dick’s novel.

History

Historical Context

After Deep Blue defeated Garry Kasparov in 1997, IBM pursued a new challenge, leading to Watson’s development. By February 2011 Watson faced Jeopardy! champions Ken Jennings and Brad Rutter, processing 500 GB of data per second and accessing 200 million pages, and won the $1 million prize, sparking commercial AI ambitions and regulatory debate in the tech sector.

1 of 3

Anthropic Dreaming

Anthropic announced Claude Managed Agents now have a “dreaming” feature that reviews past sessions.
Dreaming can run automatically or require manual approval before updating memory.
Anthropic added an outcomes loop that uses developer rubrics and a grader, raising task success by 10 points (8.4% for docx, 10.1% for pptx).
Multi-agent orchestration lets a lead agent assign specialist agents, with Netflix and Wisedocs reporting a 50% speed boost in document reviews.

1 / 3