Long-Running Agents - What Changes When Agents Run for Days

A new paradigm is emerging where AI agents run autonomously for days to weeks beyond a single chat session. This represents an evolution beyond simple question-answering to performing entire feature development or complex research.

Three Dimensions of Long-Running Execution

• Long-horizon reasoning: The ability for complex planning and execution (a domain of model quality).
• Long-running execution: Designing a harness that can withstand thousands of model calls.
• Persistent agency: The ability to maintain identity while accumulating user preferences and memory.

Three Technical Barriers Faced

• Finite context: The problem of context rot (performance degradation) that occurs before the context window fills up.
• Lack of persistent state: The 'amnesia' problem where previous work is forgotten when a session is interrupted.
• Lack of self-verification: The bias where models evaluate their own work results overly positively.

Key Design Patterns and Architecture

• Ralph loop: A practical pattern that keeps the agent itself stateless, while using the file system (JSON, txt, etc.) as external memory to record plans and progress.
• Anthropic's separated structure: Separates the model (Brain), the execution environment (Hands), and the session event log (Session) that contains all records. This secures recoverability, allowing state to be immediately reconstructed via the logs even if the sandbox is destroyed.
• Cursor's structure: Resolves the model's overconfidence problem by separating generation and evaluation through a triple structure of Planner, Worker, and Judge.