AdaptOrch Research: the paper behind the product

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

• Author: Geunbin Yu, Department of Artificial Intelligence, Korea National Open University (ORCID 0009-0006-2879-9514)
• arXiv ID: 2602.16873 (cs.MA)
• Submitted: 18 February 2026

The starting point is an observation: models from different providers now score close to each other on common benchmarks, so picking the single best model for a task buys less than it used to. The paper argues that the bigger lever is now orchestration topology, meaning how several agents are split up, run and combined.

It looks at four topologies: parallel, sequential, hierarchical and hybrid. A task is first broken into subtasks with dependencies between them, which gives a directed acyclic graph (DAG). The shape of that graph, such as how many subtasks can run side by side, how long the longest chain is and how tightly subtasks depend on each other, is used to pick the topology.

The paper lists three contributions:

1. Performance Convergence Scaling Law. A formal statement of when choosing the orchestration matters more than choosing the model.
2. Topology Routing Algorithm. Maps a task's DAG to one of the four topologies in O(|V| + |E|) time, that is, linear in the size of the graph.
3. Adaptive Synthesis Protocol. Reconciles outputs from agents that ran in parallel, with a proof that it terminates and a heuristic consistency score.

The experiments cover coding (SWE-bench), reasoning (GPQA) and retrieval-augmented generation. The author reports a 12–23% improvement over static single-topology baselines while using the same underlying models.

• Results depend on how well the task is decomposed. A poor split carries errors into every later step.
• The coupling between subtasks is estimated on a coarse scale.
• Running agents in parallel means more concurrent API calls, which can hit rate limits or budgets.
• Only three benchmarks were tested. Creative writing, long-form generation and multi-modal tasks were not.

AdaptOrch is built by Geunbin Yu, author of the AdaptOrch paper (arXiv 2602.16873). The hosted service at adaptorch.com uses the paper's routing step. When you submit a run with subtasks and dependencies, it checks that they form a valid DAG and records which topology the routing algorithm picks. The MCP tool adaptorch_route_topology lets you preview that decision without starting a run.

The product is a review layer for AI coding agents: it records what an agent ran, which checks were available and what still needs a human look. It does not inspect code or verify correctness.

The 12–23% figure comes from the paper's own experiments. It is not a measurement of the hosted product, and it says nothing about how the product will do on your repository. The claim boundary lists what we do and don't claim.

• Abstract on arXiv (2602.16873)
• Full paper (PDF)
• About AdaptOrch and ClassicMate

Want to try the product built on this work? See plans and pricing →