Anthropic just handed AI agents something they've never had before: the ability to wake up smarter than they were yesterday.
The Summary
- Anthropic launched "dreaming" at its Code with Claude conference, a system that lets AI agents review their past sessions, identify patterns in failures, and self-correct between tasks.
- Early results are dramatic: Harvey's legal AI saw 6x higher task completion rates, Wisedocs cut document review time in half, and Netflix now processes hundreds of build logs simultaneously.
- The company also moved outcomes and multi-agent orchestration from research preview to public beta, addressing the three hardest problems in production AI: accuracy, learning, and scaling complex workflows.
- Anthropic's revenue growth has exceeded internal projections as software engineers adopt Claude Code and related agent tools, with the company now pushing beyond engineering into finance and law.
The Signal
Most AI agents today are Groundhog Day machines. They wake up every morning with zero memory of yesterday's failures, doomed to repeat the same mistakes in an endless loop. Anthropic's dreaming capability breaks that cycle by running evaluations between sessions, analyzing past behavior for patterns, and helping agents establish better working methods.
The technical approach is deceptively simple. Between tasks, the system reviews old behavior, seeks patterns in what went wrong, then updates the agent's working memory with refined processes. It's not unlike how human experts internalize lessons from failed projects, except it happens automatically every time the agent goes idle.
"Getting agents to remember and learn from their previous work could make Anthropic agents more accurate."
Harvey's 6x improvement in task completion rates is the standout number here. Legal work is unforgiving. A contract review agent that hallucinates terms or misses key clauses isn't just annoying, it's a liability. That kind of jump suggests dreaming isn't just helping agents avoid obvious errors. It's helping them internalize the subtle patterns that separate mediocre legal analysis from competent work.
The other two features moving to public beta matter just as much for production deployment. Outcomes let developers define success criteria upfront, so agents know what they're optimizing for instead of guessing. Multi-agent orchestration lets teams run multiple specialized agents in parallel without creating coordination nightmares.
Netflix's use case is telling:
- Processing hundreds of build logs simultaneously
- Each agent handling specialized analysis
- Coordination happening automatically without human traffic control
That's the kind of workflow that breaks traditional sequential AI systems. You can't just throw more compute at the problem. You need agents that can work together without stepping on each other.
Anthropic is explicitly targeting sectors beyond software engineering, particularly finance and law. That makes sense. These fields have well-defined processes, clear success metrics, and tons of repetitive cognitive work that doesn't require breakthrough creativity. They're also willing to pay premium prices for tools that actually work.
The timing aligns with what Anthropic co-founder Jack Clark has been saying about self-improving AI. The company isn't waiting for some future breakthrough in artificial general intelligence. They're shipping practical self-improvement loops today, packaged as developer tools that fit into existing workflows.
The Implication
Dreaming is launching as research preview, which means developers need to apply for access. If you're building agents for high-stakes domains like legal, medical, or financial work, get in line now. The early adopter advantage here is real. Harvey didn't get 6x better overnight by accident. They got access early and had time to tune their workflows around self-improving agents.
For everyone else, watch what happens when outcomes and multi-agent orchestration hit general availability. Those features are already in public beta, which means you can start experimenting today. The real signal isn't any single feature. It's that Anthropic just made production-grade agent deployments dramatically more feasible. The gap between "cool demo" and "thing we trust with actual work" just got a lot narrower.