Claude for Long-Horizon Tasks — Lance Martin, Anthropic

The speaker illustrates how Claude models have evolved in their ability to handle autonomous work, measured by task horizon. Initially (Opus 3 era), models could manage only 10-20 minutes of autonomous work, necessitating human oversight and simpler product surfaces like autocomplete. With the rise of synchronous coding agents like Claude Code, this expanded to about an hour, enabling local execution with easier steering. The latest shift allows for significantly longer task horizons, unlocking true asynchronous agent capabilities.

Anthropic's API offerings have evolved alongside Claude’s increasing capability. The initial Messages API provided a basic prompt-response mechanism suitable for simple harnesses but lacked deployment infrastructure. The Agent SDK offered a pre-built harness for Claude Code, and most recently, Managed Agents packages both the harness and deployment infrastructure, simplifying asynchronous agent development.

To improve reliability and security for long-horizon tasks, Anthropic decoupled the 'brain' (harness) from the 'hands' (execution environments) in Managed Agents. This architecture uses a stateless harness that interacts with an append-only session log, preventing data loss if either component fails. Separating credentials from execution environments enhances security by avoiding storing sensitive information within sandboxes.

The session log in Managed Agents acts as an external context object that Claude can interrogate, enabling benefits like improved context management and compaction. This approach allows Claude to effectively manage multiple execution environments (hands) and leverage the session data for tasks such as summarizing or refining information over extended periods.

The visualization demonstrates how 'dreaming,' an offline process, corrects errors in Claude's memory. The orange line representing memory traces consistently backtracks and recovers, while the no-memory baseline remains stagnant. This correction occurs by examining prior sessions and identifying/fixing mistakes that would otherwise become permanent within Claude’s memory store.

Claude Tag is presented as a significant shift towards 'org-level harnesses,' moving beyond single-player agents. Unlike personal agent configurations that can take weeks or months to fully develop, Claude Tag offers immediate access to a well-developed harness for all organization members. This shared resource fosters collaboration and standardization across the company.

Traditionally, agents have been reactive, responding to user direction. However, asynchronous agents like Claude Tag are increasingly proactive, capable of alerting users to potentially relevant information based on organizational context. This represents a new UX paradigm where agents anticipate needs and provide timely insights rather than simply executing commands.

Benchmarks, such as Meter, demonstrate that frontier models (like Mythos or GPT-4) outperform those relying on stacking for long-horizon tasks. This suggests a fundamental difference in their capabilities for handling complex, extended reasoning processes and highlights the current advantage of these leading-edge AI systems.