wang-et-al-2025-tdag

What problem does it solve?

Prevents cascading task failures in complex, multi-step agent workflows by turning one-time plans into continuously updated, state-aware decompositions that adapt to execution results and environmental changes.

Core Features & Use Cases

• Dynamic Task Decomposition: Interleave planning and execution so each subsequent subtask is generated from actual results, not assumed outcomes.
• Just-in-Time Agent Generation: Create subagent specializations with subtask-specific tool documentation and retrieved past skills to reduce context bloat and tool misuse.
• Failure Containment & Replanning: Detect cascading failure signals, backtrack to the last valid state, and regenerate downstream subtasks to contain error propagation.
• Skill Library & Context Precision: Store and retrieve semantic skills as institutional memory and provide subtask-scoped context to keep LLM attention focused.
• Use Cases: Travel itinerary orchestration, multi-stage research pipelines, complex API-driven automations where later steps depend on earlier outcomes.

Quick Start

Replan the remaining subtasks after a failed subtask (e.g., an unavailable train booking) to adapt timelines, regenerate specialized subagents, and prevent cascading failures.