Claude Code Leak: All Upcoming Features (Allegedly)
The recent Claude Code leak exposed the internal CLI and agent harness through a published source map. While no model weights or sensitive data were revealed, the leak provided something arguably more interesting: a view into how Anthropic is building its next generation of AI agents.
What stands out is not any single feature, but a clear direction. Claude Code is evolving from a reactive assistant into a persistent, multi-agent system with memory, planning, and background execution. The following features, extracted from public analysis of the leak, represent the most meaningful signals about that shift.
KAIROS
KAIROS appears to be a persistent assistant mode designed to operate beyond a single session. It introduces long-term memory, daily logs, and the ability to act in the background even when the user is not actively interacting. This suggests a move toward proactive behavior rather than purely reactive responses. Instead of waiting for prompts, the system can maintain context over time and continue working independently. If fully realized, this would fundamentally change how developers interact with AI tools.
ULTRAPLAN
ULTRAPLAN is described as a long-running planning system that executes complex reasoning tasks in a separate environment. It can spend extended time, reportedly up to tens of minutes, exploring a problem before returning a structured plan. The user then reviews and approves execution, creating a clear separation between thinking and acting. This addresses one of the core limitations of LLMs, which struggle with long-horizon planning in a single interaction. ULTRAPLAN effectively externalizes that process into a dedicated planning phase.
Auto-Dream
Auto-Dream is a background memory consolidation system that processes past interactions into structured knowledge. It follows a pipeline of gathering, organizing, and pruning information to reduce noise and improve recall. Rather than relying on raw conversation history, the system maintains a curated representation of what matters. This reduces context size while increasing relevance in future interactions. It is conceptually similar to how human memory compresses experience over time.
Background Sessions
Background or daemon sessions allow Claude Code to run as a persistent process rather than a one-off tool. Users can start long-lived sessions, monitor them, and reconnect later to inspect results. This enables workflows where tasks continue running without active supervision. It shifts the model from a synchronous interaction pattern to an asynchronous one. In practice, this turns the assistant into a service that can operate continuously.
Coordinator Mode
Coordinator mode introduces a multi-agent system where a primary agent delegates tasks to specialized sub-agents. These sub-agents can work in parallel and return intermediate results that are aggregated into a final outcome. This allows complex problems to be broken down and solved more efficiently. It also introduces a structured approach to tool usage and reasoning. The result is a more scalable form of intelligence compared to a single-agent loop.
Computer Use
Computer use capabilities extend the agent’s ability to interact directly with the system environment. This includes executing commands, manipulating files, and potentially controlling external tools or interfaces. Some aspects of this feature are already in public preview, indicating it is closer to production than others. It represents a shift from advisory assistance to direct action. The agent becomes capable of operating software, not just suggesting how to use it.
Voice Mode
Voice mode adds a spoken interface on top of the existing CLI-based interaction. While technically straightforward, it significantly lowers the barrier to interaction. It enables more natural workflows, especially in situations where typing is inconvenient. This feature is less about capability and more about accessibility. Its importance lies in increasing usage rather than expanding core functionality.
Workflow Scripts and Templates
Workflow scripts and templates introduce reusable, higher-level task definitions. Instead of issuing multiple low-level commands, users can trigger complex workflows with a single instruction. These templates encode best practices and common patterns. Over time, they could evolve into a library of standardized operations. This is a key step toward making agents more productized and less dependent on manual prompting.
Bridge Mode
Bridge mode appears to enable interaction across environments, allowing tasks to be triggered or coordinated remotely. It suggests a model where the agent is not confined to a single local context. Instead, it can connect systems, delegate work, and synchronize results across boundaries. This is important for distributed workflows and team-based use cases. It points toward a more networked model of agent operation.
What This Actually Means
Taken together, these features point to a clear evolution. Claude Code is moving away from being a chat-based coding assistant and toward becoming a persistent, autonomous system capable of planning, remembering, and acting over time.
The most important insight from the leak is not any specific feature, but the architecture behind them. The future of AI tools is not just better models, but better systems built around them. Agents that can coordinate, run continuously, and manage their own context will define the next generation of developer tooling.
In that sense, the leak did not reveal a secret algorithm. It revealed the direction of the industry.