Anthropic is pushing the boundaries of autonomous AI agents with the introduction of "dreaming," a new system designed to let AI agents learn from their own past mistakes and improve over time. Unveiled at their second annual Code with Claude developer conference, this capability, alongside the public beta release of "outcomes" and "multi-agent orchestration," marks a significant step towards self-correcting, self-improving AI systems demanded by enterprises for production workloads.
What Happened
At its Code with Claude developer conference, Anthropic made several key announcements regarding its Claude Managed Agents platform:
- Dreaming: This novel feature allows AI agents to review their past sessions and memory stores, identify areas for improvement, and subsequently enhance their performance. Unlike conventional memory systems that simply retain context, dreaming operates at a higher level of abstraction, enabling genuine learning and self-correction.
- Outcomes & Multi-Agent Orchestration: These two features, previously in research preview, have now moved into public beta. Outcomes enable developers to define specific goals and success metrics for agents, while multi-agent orchestration facilitates the coordination of multiple agents to tackle complex, multi-step tasks efficiently.
Anthropic highlights that these three capabilities directly address the most challenging aspects of deploying AI agents at scale: maintaining accuracy, fostering continuous learning, and preventing bottlenecks in intricate workflows.
Early adopters are already seeing substantial benefits:
- Legal AI company Harvey reported a roughly 6x increase in task completion rates after implementing dreaming.
- Medical document review company Wisedocs cut its document review time by 50% using outcomes.
- Netflix is leveraging multi-agent orchestration to process logs from hundreds of builds simultaneously.
The company also disclosed explosive growth, with CEO Dario Amodei stating Anthropic saw 80x annualized growth in revenue and usage in Q1 2026, far exceeding its projected 10x growth. API volume is up nearly 70x year over year, with developers spending an average of 20 hours per week using Claude Code.
Image 1: anthropic growth graph: image omitted due to site embedding policy; open the original article (VentureBeat) (opens in a new tab) to view it. Photo/source: VentureBeat (https://venturebeat.com/technology/anthropic-introduces-dreaming-a-system-that-lets-ai-agents-learn-from-their-own-mistakes (opens in a new tab))
How Anthropic's dreaming feature teaches AI agents to learn from their own history
Dreaming distinguishes itself from basic agent memory, which primarily allows Claude to retain preferences and context within or across individual sessions. Instead, dreaming is a more profound, scheduled process. It systematically reviews an agent's historical sessions and existing memory stores, identifying patterns, successes, and failures. This retrospective analysis enables the agent to abstract lessons from its experiences, allowing it to adapt its strategies and improve its future decision-making without explicit retraining or constant human intervention.
Why It Matters
For developers and IT leaders, Anthropic's new features, particularly "dreaming," are a game-changer for several reasons:
- Enhanced Reliability and Trust: The ability for AI agents to self-correct and improve directly addresses a major enterprise concern: the reliability of autonomous systems. If agents can learn from their errors, they become more robust and trustworthy for critical production tasks, reducing the need for constant human oversight and intervention.
- Scalability for Complex Workflows: Multi-agent orchestration directly tackles the bottleneck issue when dealing with intricate, multi-step processes. Developers can now design systems where specialized agents collaborate seamlessly, mirroring human team dynamics and enabling the automation of previously unmanageable workloads.
- Reduced Development and Maintenance Overhead: Agents that learn and improve autonomously reduce the need for iterative fine-tuning and retraining. This frees up developer resources, accelerates deployment cycles, and lowers the long-term maintenance costs associated with AI applications.
- Driving Broader AI Adoption: The practical successes reported by early adopters like Harvey and Wisedocs demonstrate that these aren't just theoretical advancements. Improved task completion rates and reduced processing times directly translate into tangible ROI, providing a strong incentive for broader enterprise adoption of AI agent solutions.
- Addressing Compute Demands: While Anthropic's rapid growth indicates immense demand, CEO Amodei's comment about compute difficulties underscores the ongoing challenge of scaling AI infrastructure. Features that make agents more efficient and self-sufficient could, in the long run, help optimize resource utilization.
What To Watch
The introduction of "dreaming" is a significant step towards truly autonomous and adaptive AI agents. Developers should closely monitor how this feature evolves, particularly in its ability to generalize learnings across diverse tasks and environments. Will we see agents able to abstract lessons from one domain and apply them effectively to another?
Keep an eye on the ecosystem of tools and best practices that emerge around multi-agent orchestration. The ability to coordinate complex AI workflows will require new patterns and frameworks for development, monitoring, and debugging. As these capabilities mature, we can expect to see a new generation of highly intelligent and resilient AI applications capable of tackling challenges previously thought too complex for automation. The continued exponential growth of Anthropic also indicates that the demand for advanced AI capabilities is skyrocketing, pushing the boundaries of what's possible with agentic AI.