Anthropic reveals for the first time how the next-generation Claude is built! User complaints directly fed into the model, even AI's 'dreams' are being trained

TL;DR

• Anthropic has shifted its model development process from pure research to a productized engineering workflow, defining specific capabilities, target scenarios, and defect fixes for each new Claude iteration before training begins.
• The next generation of Claude is evolving into a long-running agent system featuring "Adaptive Thinking," which allows the model to autonomously determine when to engage in deep reasoning based on the complexity of the query.
• Anthropic has introduced a "dreaming" mechanism for managed agents, where the system reviews, consolidates, and clears conflicting memory data during idle periods, analogous to human memory reconsolidation during sleep.
• Internal workflow analysis reveals that while coding efficiency has increased by up to 100x, the primary bottleneck has shifted from code generation to organizational coordination, strategic decision-making, and cross-team communication.
• The company is systematically training model "personality" and judgment boundaries to ensure trustworthiness in autonomous agents, with consciousness research also formally added to the internal research agenda.

Why It Matters

This signals a fundamental maturity shift in the AI industry, moving away from benchmark chasing toward the deployment of reliable, long-term autonomous agents. For practitioners, the focus on structured memory consolidation ("dreaming") and adaptive resource allocation (Adaptive Thinking) provides a blueprint for building agents that can operate continuously without degrading context quality. Additionally, the insight that organizational coordination, not code, is now the bottleneck highlights the changing value of human expertise in AI-driven workflows.

Key Data

• Efficiency Improvement: Code-level efficiency gains are estimated at 100x, whereas organizational coordination has not achieved similar acceleration.
• Development Speed: Prototypes and MVPs that previously took 2-3 weeks can now be built in one day.
• Decision Latency: Product managers can now retrieve data and log analysis in 10 minutes using Claude Code, replacing a process that previously took days of waiting for data science teams.
• User Base: Millions of users interact with Claude daily, providing large-scale feedback for iterative improvement.
• Research Focus: Consciousness has been formally included as an internal research topic, with dedicated staff studying whether Claude can become a conscious actor.

Technical Details

• Adaptive Thinking: Unlike "Extended Thinking" (which forces deep reasoning upon activation), this feature allows the model to autonomously decide when to enter deep reasoning modes. It allocates computational resources based on problem complexity, requiring sufficient user context to make accurate judgments about when to think deeper.
• The "Dreaming" Mechanism: A background process active during agent idle time that traverses memory to identify conflicts, delete invalid information, compress context, and reconstruct user profiles. This is explicitly designed to mimic the memory reconsolidation process in human sleep, ensuring long-term agents maintain clean, coherent state.
• Productized Model Training: Each model generation is treated as a product with defined specifications, target capabilities, and defect lists derived from customer feedback. Product managers are involved from the conceptual phase through to release, ensuring alignment between technical training decisions and user scenarios.
• Automated Feedback Loop: Claude is used internally to cluster user feedback, identify core themes, and convert qualitative issues into quantitative evaluation items (evals), enabling rapid diagnosis of model behavior problems.

Industry Insight

• Shift to Coordination as the Bottleneck: As coding costs approach zero, AI leaders must focus on strategic alignment, irreversible decision-making ("one-way doors"), and human communication. The value of product managers will lie in orchestrating these human-centric tasks rather than estimating technical effort.
• Memory as a Core Feature: The "dreaming" mechanism suggests that the next competitive frontier for agents is not just raw reasoning power, but the ability to maintain stable, non-conflicting long-term memory. Companies building agents should prioritize background context maintenance and conflict resolution.
• Personality and Trust Engineering: Systematically training "personality" and judgment boundaries is becoming a critical component of agent safety. As agents run for longer durations, their ability to refuse bad requests or challenge user assumptions becomes a key trust metric, arguably more important than raw code capability.

zing token usage and response latency based on the difficulty of the task.

Q: What is the practical purpose of the "dreaming" mechanism?
A: The "dreaming" mechanism allows agents to perform background memory maintenance during idle times. It reviews existing memories to find contradictions, removes invalid data, and compresses context, ensuring the agent has a clean and accurate state for future interactions, similar to how humans consolidate memories during sleep.

Q: Why does Anthropic emphasize that the bottleneck has shifted to organizational coordination?
A: Because AI has made code generation and prototyping extremely fast (down to one day for MVPs), the time spent writing code is no longer the limiting factor. The slowest part of the development cycle is now making strategic decisions, coordinating across teams, and communicating with stakeholders, which AI cannot yet fully automate.