PaperlayerView exact revision on arXiv

Source-linked AI summary

SkillOrchestra: Learning to Route Agents via Skill Transfer

Jiayu Wang, Yifei Ming, Zixuan Ke, Shafiq Joty, Aws Albarghouthi, Frederic Sala

arXiv:2602.19672v1cs.AIcs.LG

TL;DR

Existing routers make coarse query-level decisions, while RL orchestrators can be costly to adapt and prone to routing collapse. SkillOrchestra learns a reusable Skill Handbook that models skill-conditioned agent competence and cost, then selects agents for current interaction demands. Across ten benchmarks, it outperforms RL-based orchestrators by up to 22.5% with substantially lower learning cost and transfers across orchestrator backbones without retraining.

  • Problem

    Existing routing methods overlook evolving skill requirements in multi-turn tasks, while RL-trained orchestrators are expensive to adapt and prone to routing collapse.

  • Method

    SkillOrchestra learns a reusable Skill Handbook from execution experience and uses state-conditioned mode and skill information to select agents under a performance-cost trade-off.

  • Results

    Across ten benchmarks, SkillOrchestra improves accuracy, efficiency, and routing stability, achieving up to 22.5% absolute improvement over SoTA RL-trained orchestrators with 700× and 300× lower learning cost than Router-R1 and ToolOrchestra.

  • Takeaways & Limitations

    The Skill Handbook transfers across orchestrator backbones and updated model pools without retraining, supporting scalable orchestration as agent pools evolve.

  • Takeaways & Limitations

    Fine-grained handbooks can degrade performance for lower-capacity orchestrators that misidentify the active subskill, introducing routing bias.

Abstract

from arXiv · show

Compound AI systems promise capabilities beyond those of individual models, yet their success depends critically on effective orchestration. Existing routing approaches face two limitations: (1) input-level routers make coarse query-level decisions that ignore evolving task requirements; (2) RL-trained orchestrators are expensive to adapt and often suffer from routing collapse, repeatedly invoking one strong but costly option in multi-turn scenarios. We introduce SkillOrchestra, a framework for skill-aware orchestration. Instead of directly learning a routing policy end-to-end, SkillOrchestra learns fine-grained skills from execution experience and models agent-specific competence and cost under those skills. At deployment, the orchestrator infers the skill demands of the current interaction and selects agents that best satisfy them under an explicit performance-cost trade-off. Extensive experiments across ten benchmarks demonstrate that SkillOrchestra outperforms SoTA RL-based orchestrators by up to 22.5% with 700x and 300x learning cost reduction compared to Router-R1 and ToolOrchestra, respectively. These results show that explicit skill modeling enables scalable, interpretable, and sample-efficient orchestration, offering a principled alternative to data-intensive RL-based approaches. The code is available at: https://github.com/jiayuww/SkillOrchestra.

1. Introduction

SkillOrchestra addresses the limits of coarse query-level routing and costly RL orchestration by modeling reusable skills, agent competence, and cost. Across ten benchmarks, it improves performance-cost trade-offs, routing stability, and transferability.

  • Motivation: Single-shot routers miss changing capability requirements across multi-turn interactions and operational modes such as web search and coding.Routing should be conditioned on the current interaction state rather than treating the entire query as one decision.
  • Skill-aware orchestration: SkillOrchestra learns a reusable Skill Handbook from execution experience, combining mode-level insights, fine-grained skills, and skill-conditioned agent profiles.Agent profiles capture performance, cost characteristics, and practical usage insights.
  • Skill-aware orchestration: At deployment, the orchestrator selects an operational mode and then chooses an agent that satisfies the required skills under an explicit performance-cost trade-off.This enables state-conditioned and fine-grained orchestration rather than direct end-to-end routing-policy optimization.
  • Benefits: SkillOrchestra promotes balanced routing, mitigates routing collapse, and transfers orchestration knowledge across orchestrator backbones and updated model pools.The handbook decouples orchestration knowledge from router parameters.
  • Results: Across ten benchmarks, SkillOrchestra outperforms strong RL-trained orchestrators by up to 22.5% while reducing learning cost 700× versus Router-R1 and 300× versus ToolOrchestra.It is evaluated in both multi-turn model routing and full agent orchestration settings.

2. Related Works

Prior work includes heuristic and discriminative model routers and newer RL-based sequential routing systems. SkillOrchestra instead introduces skill as an intermediate abstraction and uses a reusable handbook for orchestration.

  • Model Routing: Earlier model-routing methods use heuristics, cascades, or discriminative mappings from query features to model choices.These approaches primarily target selecting a model from a pool while balancing performance and inference cost.
  • RL-based Routing and Orchestration: RL-based systems formulate routing as a sequential decision process and optimize performance-cost trade-offs through trajectory-level rewards.Router-R1 and ToolOrchestra interleave reasoning and routing for multi-step decisions.
  • Skill-aware orchestration: SkillOrchestra’s overview contrasts static model routing and implicit end-to-end orchestration with explicit skill-level capability modeling and balanced agent utilization.The framework uses a reusable Skill Handbook to support extensibility.
  • RL-based Routing and Orchestration: RL-based approaches face high training cost, limited adaptability to new pools or tasks, and routing collapse toward repeatedly invoking one strong but expensive model.SkillOrchestra addresses these challenges through a reusable Skill Handbook and an intermediate skill abstraction.

3. Preliminaries

The paper formalizes agent orchestration as state-dependent selection of operational modes and agents in multi-step workflows, optimizing performance-cost tradeoffs. SkillOrchestra replaces direct end-to-end routing optimization with a reusable Skill Handbook that links task demands to agent capabilities.

  • Agent Orchestration: At each turn, the orchestrator selects an operational mode and an agent capable of executing it, producing traces and observations that define the evolving trajectory.Modes specify what operation is required, while agents execute within mode-specific tool constraints.
  • Problem Formulation: The orchestration objective maximizes expected trajectory reward while minimizing cumulative agent execution cost through an explicit tradeoff.Cost can include token usage and latency, and the policy separates mode selection from agent routing.
  • Problem Formulation: Traditional model routing is a special case with one timestep, one answer mode, no external tools, and one model selected for the final response.This formulation captures routing as a single-step policy conditioned only on the user query.
  • SkillOrchestra: Existing RL approaches directly finetune orchestrator parameters toward a routing policy, whereas SkillOrchestra learns a structured handbook from execution experience.The handbook contains execution insights, fine-grained skills, and agent profiles describing competence and cost under those skills.
  • SkillOrchestra: SkillOrchestra’s handbook supports state-conditioned, fine-grained orchestration, more stable routing, and reuse across orchestrator backbones and updated model pools.The framework is designed to mitigate routing collapse while decoupling orchestration knowledge from router parameters.

4. SkillOrchestra

SkillOrchestra uses a structured Skill Handbook to guide mode selection and skill-grounded agent routing. It adapts handbook granularity to orchestrator capacity and selects agents by balancing skill-conditioned competence against execution cost.

  • Skill Handbook: Skills provide an intermediate layer between operational modes and agents, decoupling capability requirements from agent identity.The handbook can represent broad skills such as data_processing and more specialized subskills such as symbolic_logic.
  • Skill Handbook: The Skill Handbook organizes mode-level insights, reusable skills, and agent profiles into a structured source of orchestration knowledge.Profiles store skill-conditioned competence, routing signals, and execution-cost information.
  • Runtime Orchestration: The mode–skill index restricts runtime routing to skills associated with the selected mode rather than searching the full skill space.This mapping connects each operational mode to its mode-consistent skill set.
  • Runtime Orchestration: At inference time, the orchestrator retrieves a task-conditioned handbook, selects an operational mode, identifies active skills, and routes to an agent using competence and cost.Mode selection uses interaction state and routing metadata; agent selection aggregates competence estimates over relevant skills.
  • Handbook Selection: Fine-grained skill structure can harm lower-capacity orchestrators when they misidentify the active subskill, so handbook granularity is selected through validation for the target orchestrator.Coarser skills can reduce sensitivity to misidentification and produce more stable routing under such conditions.
  • Handbook Selection: Handbook subsets are evaluated on held-out trajectories using task success minus cost, with λ controlling the performance-cost trade-off.The objective evaluates complete trajectories and selects a Pareto-frontier handbook for the target orchestrator.

5. Experiments

Experiments evaluate SkillOrchestra across model routing and full agent orchestration, testing effectiveness, efficiency, routing behavior, transferability, and handbook components. Across these settings, it improves performance-cost trade-offs, mitigates routing collapse, transfers across orchestrator backbones, and benefits from refined, fine-grained skills.

  • Experimental setup: SkillOrchestra is evaluated across ten diverse benchmarks against heuristic, discriminative, and RL-based methods.The experiments cover model routing and full agent orchestration settings.
  • Effectiveness: SkillOrchestra reaches 47.4 EM versus Router-R1’s 41.6 EM, while SkillOrchestra+ reaches 51.6 EM.On multihop tasks, Musique improves from 13.8 to 18.2 to 20.6, and Bamboogle from 51.2 to 58.4 to 63.2.
  • Efficiency: 22.5 percentage points is the maximum reported accuracy improvement over Router-R1 in math reasoning, while inference cost decreases by about 2×.The comparison is reported in Figure 5.
  • Routing behavior: Router-R1 selects LLaMA-3.1-70B for 98.02% of calls, whereas SkillOrchestra distributes calls across models according to capability differences.SkillOrchestra’s example distribution assigns 44.53% to Mixtral-8×22B, 25.99% to Qwen2.5-7B, 15.38% to LLaMA-3.1-70B, and 11.50% to Qwen2.5-3B.
  • Transferability: A Skill Handbook learned with Qwen2.5-3B improves other orchestrator backbones without additional handbook training.Across three general QA datasets, gains include +24.3 for Qwen2.5-7B, +22.5 for Llama3.1-8B, and +23.3 for Mistral-7B.
  • Full agent orchestration: SkillOrchestra achieves 84.3% accuracy at $72.7 total cost in full agent orchestration, versus ToolOrchestra’s 76.3% and $92.7.This corresponds to an 8.0-point accuracy improvement and a 21.6% cost reduction.
  • Component contribution: Removing the Skill Handbook drops accuracy from 85.0% to 71.0% and raises cost from 9.3 to 122.9.Refinement lowers cost from 5.5 to 3.4 while maintaining similar accuracy, and fine-grained skills support better decisions.

6. Conclusion

SkillOrchestra reframes multi-turn routing as skill-grounded decision making and uses a Skill Handbook for state-aware, competence-aware performance-cost optimization. Across model routing and agent orchestration, it reports stronger performance at lower cost, transferable orchestration knowledge, and deployment as model pools evolve.

  • Conclusion: SkillOrchestra reframes multi-turn routing as skill-grounded decision making through a learned Skill Handbook.The framework makes state-aware and competence-aware decisions that explicitly optimize the performance-cost trade-off.
  • Conclusion: Across model routing and agent orchestration, SkillOrchestra achieves superior performance with significantly lower cost than competitive baselines.The conclusion also reports transfer across orchestrator backbones without retraining.
  • Conclusion: The transferable Skill Handbook supports scalable deployment as model pools evolve.The reported transfer occurs across orchestrator backbones without retraining.

A.1. Experimental Details for Model Routing

The model-routing experiments use a controlled Router-R1 protocol with Qwen2.5-3B as orchestrator and a six-model pool. They evaluate Exact Match and total completion cost under a low-data training regime.

  • Setup: The model pool contains Qwen2.5-7B, LLaMA-3.1-8B, LLaMA-3.1-70B, Mistral-7B, Mixtral-8x22B, and Gemma-2-27B.Qwen2.5-3B serves as the orchestrator under the Router-R1 evaluation protocol.
  • Routing protocol: Routing uses search mode for subtasks and answer mode for aggregating intermediate results into the final answer.The maximum interaction horizon is four turns.
  • Evaluation: Performance is measured with Exact Match, while efficiency is measured using total completion cost.SkillOrchestra training selects fewer than 50 samples per dataset by default, plus the same number for validation and handbook retrieval.

A.2. Experimental Details for Agent Orchestration

The agent-orchestration experiments follow ToolOrchestra’s protocol across search, code, and answer modes with a 50-turn horizon. They measure answer accuracy using GPT-5-mini as judge and total system cost in USD.

  • Operational modes: Search mode permits WebSearch and LocalSearch, with GPT-5, GPT-5-mini, and Qwen3-32B as search models.LocalSearch uses a FAISS index built with Qwen3-Embedding-8B.
  • Evaluation: The maximum interaction horizon is 50 turns, with accuracy judged by GPT-5-mini and total system cost measured in USD.The setup follows ToolOrchestra for controlled comparison.
  • Operational modes: Answer mode uses no external tools and permits GPT-5, GPT-5-mini, Llama-3.3-70B-Instruct, Qwen3-32B, Qwen2.5-Math-72B, and Qwen2.5-Math-7B.Valid answer agents are model-only compositions.

B. Skill-Grounded Agent Routing Algorithm Pseudocode

The section presents Algorithm 1 for skill-grounded agent routing and points to Figure 3 for a concrete deployment illustration.

  • Algorithm 1 presents the procedure for skill-grounded agent routing by the orchestrator.
  • The pseudocode section connects the routing algorithm with its deployment illustration.
  • Figure 3 provides a concrete illustration of the deployment process.

C. A Closer Look at Model Selection: SkillOrchestra vs. ToolOrchestra

SkillOrchestra allocates models by operational mode and skill demands rather than simply minimizing call counts, avoiding ToolOrchestra’s concentrated routing patterns.

  • SkillOrchestra reduces cost through smarter allocation of models across operational modes rather than fewer calls.
  • Routing procedure: The routing algorithm retrieves active skills, estimates competence, subtracts mode-specific cost, executes the selected agent, and updates state.Its utility is defined as competence minus a cost-weighted mode-specific cost.
  • Answer mode: In answer mode, ToolOrchestra routes 97.9% of calls to GPT-5, indicating another concentrated routing pattern.
  • Answer mode: SkillOrchestra uses GPT-5 for 58.4% of answer calls, with GPT-5-mini at 10.0% and the remainder handled by cheaper or specialized models.

D. Demonstrations of Skill-Aware Orchestration

The demonstrations show that skill-aware orchestration uses a Skill Handbook and model-specific skill profiles to select capabilities across diverse tasks, including cases where no external call is needed.

  • Demonstrations: The Skill Handbook supports an example with two external search-model calls that produces a correct result.
  • Math demonstrations: Qwen2.5-7B-Instruct achieves 25% success, including 100% on 3D vector geometry and harmonic-denominator versus LCM problems.
  • Demonstrations: An AMC example shows the skill-based router recognizing that no external model is needed and answering autonomously.
  • Math demonstrations: LLaMA-3.1-70B-Instruct achieves 30% success on the tested math batch, with full success on sphere-plane sections and base-conversion equations.
  • Math demonstrations: The remaining model profiles show varied or weak skill coverage, including 5% for Mistral-7B-Instruct and 0% for Gemma-2-27B-Instruct on one math batch.
  • Knowledge demonstrations: Across the tested entity and relation tasks, Gemma-2-27B-Instruct achieves 48% success, while Mixtral-8x22B-Instruct achieves 45%.
Loading 2602.19672v1…