Why Your Agent Can't Follow a Plan (And How to Fix It)

The article discusses common failure modes in production agents and proposes a workflow architecture using an explicit and executable TaskTree to address issues like implicit task structure, lack of failure isolation, and no resumability.

Details

The author has been building production agents and identified three common failure modes: implicit task structure, no failure isolation, and no resumability. To address these issues, the article introduces the TaskTree abstraction, which represents the workflow plan as an explicit and executable data structure. The TaskTree contains tasks with metadata like dependencies, status, and results. The agent simply calls get_ready_tasks() to determine the next steps to execute, rather than trying to plan its own actions. The initial TaskTree is generated using an LLM-powered decomposition process, where the language model breaks down the high-level goal into concrete, executable subtasks. This separation of planning and execution allows the agent to focus on reliable task completion without getting lost in its own context.