A new paper proposes letting AI agents build their own tool hierarchy instead of juggling every tool in one long reasoning chain.
The method, called tool-aware recursive decomposition or TaReD, targets a specific failure mode in long-horizon agent tasks: a single chain of reasoning and action that grows so long its own history buries the dependencies between steps, letting early planning errors snowball. Rather than loading every tool description into context upfront, which burns tokens and makes the right tool harder to find in a crowded list, TaReD groups tools by function into a capability hierarchy and has the agent pull in tools on demand. The agent then recursively breaks the task into a tree of subtasks, with each level of the tree matched to the capabilities needed at that stage. On complex real-world tasks, the researchers report end-to-end success rates up to 40 percentage points higher than their baseline agents, and they have posted the implementation on GitHub.
Tool bloat is becoming a real bottleneck as agents get plugged into dozens of APIs and MCP servers at once, and most agent designs still lean on a flat ReAct-style loop built for a handful of tools, not hundreds. TaReD's bet is that the fix is structural: organize the tool space like an org chart, not just a bigger context window or a smarter single-shot plan.
The 40-point jump is measured against the authors' own baselines on their own task set, not an independently run benchmark, so treat it as a promising lab result until someone else tries to reproduce it.