Anthropic’s latest leap in autonomous artificial intelligence (AI) capabilities now includes a “dreaming” feature that lets AI agents analyze past performance and self-correct, a move directly aimed at winning over enterprises that remain wary of deploying autonomous systems at scale.
The announcement comes amid what CEO Dario Amodei described as “extraordinary momentum.” Amodei revealed that Anthropic’s growth has obliterated internal forecasts, recording an 80-times annualized increase in revenue and usage in the first quarter of 2026, dramatically higher than the 10-times growth the company anticipated.
The centerpiece of the update, “dreaming,” moves beyond simple memory.
Standard AI memory allows a model to remember a user’s name or a previous prompt, but dreaming is a scheduled background process. It reviews an agent’s past sessions, identifies recurring errors, and extracts successful patterns to create structured playbooks.
“We’re not changing the model weights,” said Alex Albert, Anthropic’s research product lead. Instead, the agent writes plain-text notes to its “future self,” making the learning process entirely auditable by humans. Early adopters are already seeing radical efficiency gains; legal AI firm Harvey reported a six-times increase in task completion rates after implementing the feature.
Alongside dreaming, Anthropic moved two critical features from research preview into public beta. Outcomes let developers set a success rubric such as brand voice and technical standards). A separate grader agent then evaluates the primary agent’s work in a fresh context window, forcing iterations until the standard is met.
Secondly, multi-agent orchestration enables a lead agent to delegate complex tasks to specialized sub-agents. Netflix Inc. is reportedly using this to process logs from hundreds of builds simultaneously.
“Agent platforms are competing to own the memory layer, and Anthropic’s dreaming, outcomes, and multi-agent orchestration release moves Claude Managed Agents from API to control plane. Immutable input stores and reviewable consolidated memory are the right enterprise signal,” said Mitch Ashley, vice president and practice lead for Software Lifecycle Engineering at The Futurum Group. “Scaffolding is the biggest gap. Memory that mutates between runs needs lineage, poisoning detection, and drift telemetry the current controls do not yet expose. Background consolidation generates compute disconnected from user-visible work, and enterprises will require runtime cost governance before granting agents the autonomy this architecture intends to enable.”
Anthropic Chief Product Officer Ami Vora noted that while AI capabilities are advancing exponentially, organizational adoption has remained linear. These tools are designed to bridge that capability gap by allowing agents to run for hours or days without human intervention.
The massive 80-times growth has not been without growing pains. Amodei admitted the company has faced “difficulties with compute” due to the unexpected demand. To combat this, Anthropic announced a partnership with SpaceX to utilize the full capacity of its Colossus data center, alongside a significant increase in API rate limits for enterprise users.
The conference also highlighted massive industrial shifts. Latin American e-commerce giant Mercado Libre revealed that 23,000 of its engineers are now using Claude Code, with the goal of reaching 90% autonomous coding by Q3 2026.
Amodei revisited his bold prediction that 2026 would see the rise of the first billion-dollar company run by a single person.
“Hasn’t quite happened yet,” Amodei said, “but we’ve got seven more months.”
With the launch of self-improving agents, that timeline may be closer than ever.
