Here is a sentence that would have sounded insane two years ago: Anthropic is giving its AI agents the ability to dream.

At its second annual Code with Claude developer conference in San Francisco on May 7, Anthropic unveiled "dreaming" for Claude Managed Agents. It is a scheduled background process that reviews an agent's past sessions, extracts patterns across them, identifies recurring mistakes, and curates learnings into structured playbooks that future sessions can reference. The agents are, in a very literal sense, learning from their own experience while nobody is watching.

The results are already striking. Legal AI company Harvey saw task completion rates increase roughly 6x after implementing dreaming. Medical document review company Wisedocs cut its document review time by 50% using Anthropic's new outcomes feature. And Netflix is processing logs from hundreds of builds simultaneously using multi-agent orchestration.

How Dreaming Actually Works

The critical distinction: dreaming does not modify Claude's underlying model weights. "We're not changing the model itself through dreaming," Alex Albert, who leads research product management at Anthropic, told VentureBeat. Instead, the agent writes learnings as plain-text notes and structured playbooks that future sessions can reference. Everything is observable and auditable by humans.

Albert described the concept as analogous to how people within organizations create SOPs after working through a task. "A very similar thing is happening with dreaming. Instead of you manually creating the skill from your experience working with Claude, the model is doing it, so it has that same context for a future session."

This is not memory (which Anthropic already launched earlier this year for retaining preferences within sessions). This is meta-learning. The agent reviews sessions it never participated in, finds patterns no single session could see, and synthesizes those patterns into institutional knowledge. It surfaces recurring mistakes, workflows that multiple agents independently converge on, and team-wide preferences.

The Live Demo Was Convincing

Anthropic demonstrated all three features live on stage using a fictional aerospace startup that needed to autonomously land drones on the moon. They configured a multi-agent system with three specialists (commander, detector, navigator), ran initial simulations that produced imperfect results, triggered a dreaming session, and ran new simulations the next morning. The results improved meaningfully on previously underperforming scenarios.

"All we had to do was just have Caitlin press a button," said Angela Jiang, Head of Product for the Claude Platform. "All dreaming." The implication is clear: enterprise customers can set this up, go home, and come back to meaningfully better agents in the morning.

The Growth Numbers Are Absurd

CEO Dario Amodei disclosed during a fireside chat that the company's growth has outpaced even its own aggressive internal projections. In Q1 2026, Anthropic saw what Amodei described as 80x annualized growth in revenue and usage. Not 80%. 80x. The company had planned for 10x annual growth, and instead got eight times that.

API volume on the Claude platform is up nearly 70x year over year. The average developer using Claude Code now spends 20 hours per week with the tool. "We tried to plan very well for a world of 10x growth per year," Amodei said. "And yet we saw 80x. And so that is the reason we have had difficulties with compute." That is not a humble brag. That is a capacity crisis framed as a victory lap.

Why This Matters More Than Another Model Release

Dreaming, combined with outcomes (now in public beta for defining success rubrics) and multi-agent orchestration (also public beta for splitting complex tasks across specialists), forms what Anthropic calls a "continuous improvement loop." Agents run, get graded, dream about what went wrong, and perform better next time. This is the self-improving enterprise AI system that every CTO has been asking for before trusting agents with production workloads.

The Atlantic published an analysis this week arguing that AI might not be a bubble after all, pointing to Claude Code as the product that crossed the threshold from "interesting gadget" to "life-changing technology." Enterprise AI spending on tool subscriptions has gone from 25% of US businesses at the start of 2025 to over 50% today. Goldman Sachs found companies are overrunning their initial AI budgets "by orders of magnitude."

Anthropic is not just building smarter models. It is building models that get smarter on their own. That is a fundamentally different value proposition, and enterprise buyers are responding accordingly.