research(orchestration): AdaptOrch dynamic DAG topology selection — parallel/sequential/hierarchical/hybrid switching (arXiv:2602.16873) #2434
Description
Summary
arXiv:2602.16873 — AdaptOrch: Task-Adaptive Multi-Agent Orchestration in the Era of LLM Performance Convergence (February 2026)
Technique
AdaptOrch presents a framework for dynamically selecting among four canonical DAG topologies (parallel, sequential, hierarchical, hybrid) based on task dependency graphs. Uses a bandit-style approach to learn which topology performs best for which task characteristics.
Related: MoMA (Mixture of Models and Agents) — generalizes routing across both models and agents, treating orchestration strategy selection as a bandit problem.
Applicability to Zeph
Zeph's LlmPlanner currently generates a single fixed DAG topology per plan. AdaptOrch suggests:
- Classifying the task dependency structure before planning
- Selecting topology family (parallel for independent subtasks, sequential for ordered, hierarchical for delegated subplans)
- Potentially learning which topology family works best for which task type
Complements #2424 (utility-guided orchestration) and #2425 (cascade-aware routing).
Priority
P3 — useful orchestration improvement, not blocking.
Source
arXiv:2602.16873 (February 2026)