wang-et-al-2025-tdag

Prevent cascading failures in multi-step agent workflows by using dynamic task decomposition, which detects failure signals, backtracks to the last valid state, and regenerates downstream subtasks to contain error propagation.

Dynamic task decomposition for adaptive replanning is a technique that interleaves planning and execution, generating each subsequent subtask from actual execution results rather than assumed outcomes to adapt to environmental changes.

Reduce context bloat and tool misuse during subagent generation by applying just-in-time specialization, which provides subtask-specific tool documentation and subtask-scoped context to keep LLM attention focused.

Yes, you can use dynamic replanning for complex travel itinerary orchestration. It adapts timelines by replanning remaining subtasks after a failed step, such as an unavailable train booking, to prevent workflow disruption.

Semantic skill retrieval supports multi-stage research pipelines by storing and retrieving past skills as institutional memory, ensuring subtask-scoped context is available to maintain focus and prevent errors in interdependent steps.

You should not use just-in-time agent generation for simple, linear tasks that lack interdependencies, as the overhead of dynamic replanning, backtracking, and subtask-scoped context provides no benefit without complex, multi-step execution requirements.