Back to AI Research

AI Research

HTN-Orchestrate separates global planning from local agent execution

Key Takeaways

  • A hierarchical task network tracks dependencies while local orchestrators dispatch workers, allowing decomposition to change as evidence arrives.
  • ## Structure without fixing every step in advance
  • HTN-Orchestrate addresses the tension between flexible agent dispatch and a coherent plan.
  • A dispatcher can react to local needs but lose track of dependencies.
  • A predefined workflow supplies structure, but commits to a decomposition before the task has produced the information needed to choose later steps.

Structure without fixing every step in advance

HTN-Orchestrate addresses the tension between flexible agent dispatch and a coherent plan. A dispatcher can react to local needs but lose track of dependencies. A predefined workflow supplies structure, but commits to a decomposition before the task has produced the information needed to choose later steps.
The researchers propose a two-layer system built around a hierarchical task network. A Planner maintains the subtask hierarchy and its dependencies. Local Orchestrators schedule workers for leaf subtasks, separating responsibility for the overall plan from the execution of individual pieces.

Let results inform the next decomposition

The Planner can defer decomposition until prerequisite results are available. Results propagate upward under AND semantics, meaning the framework does not treat one successful child as sufficient for a parent that requires several parts. When the current decomposition cannot resolve the task, accumulated findings inform a rebuilt network.
The design therefore makes relationships between tasks explicit while retaining a way to revise the plan. That is a specific orchestration approach, rather than a claim that adding more agents automatically improves every workflow. Its usefulness depends on whether the task's dependencies and completion conditions are represented well.

Compare like-for-like results

The authors evaluate five benchmarks spanning web research, scientific reasoning and research workflows, using the same base model for their comparisons. They report 89.73% accuracy on BrowseComp and 73.44% on SUPERChem, exceeding Agent Team by 2.45 and 4.69 percentage points respectively.
Those figures describe the paper's benchmark configuration. They do not establish the cost, speed or reliability of an arbitrary production workflow. Franklin's reading is that a useful reproduction should inspect where planning was deferred, when the network was rebuilt and which required subtasks were actually completed. Those observations would explain whether the hierarchy improved coordination, rather than merely recording a higher final score.
Read the HTN-Orchestrate paper.

Comments