Paper

AdaptOrch: Task-Adaptive Multi-Agent Orchestration in the Era of LLM Performance Convergence

• arXiv: https://arxiv.org/abs/2602.16873 (submitted 18 February 2026)

Summary

AdaptOrch formalizes a "Performance Convergence Scaling Law": as frontier models converge in capability, orchestration topology (parallel, sequential, hierarchical, hybrid) now dominates system-level performance more than individual model selection. It introduces a DAG-based Topology Routing Algorithm that selects the optimal multi-agent pattern in O(|V|+|E|) time, showing 12–23% gains over static single-topology baselines on SWE-bench, GPQA, and RAG tasks.

Relevance to Zeph

The existing TaskGraph/DagScheduler/AgentRouter stack in zeph-core/zeph-orchestration uses a fixed orchestration pattern (tasks are assigned roles sequentially by LlmPlanner). AdaptOrch provides:

• Theoretical basis for making topology selection dynamic based on task decomposition characteristics
• A concrete routing algorithm that could extend AgentRouter::route() to select between parallel fan-out, sequential chain, and hierarchical delegation patterns at plan time
• Benchmark evidence that this matters more than model selection post-convergence — directly relevant given Zeph's multi-model design

Distinct from VMAO (#2252), which focuses on replanning after task failure, not initial topology selection.

Priority

P2 — extends an existing orchestration subsystem with a principled topology selection layer.