MAS-Orchestra: LLM Multi-Agent Orchestration

• MAS-Orchestra is a multi-agent system framework that employs function-calling reinforcement learning to design complete agent ensembles at training time.
• It controls orchestration capacity through the Degree of MAS (DoM), optimizing candidate policies with GRPO and KL regularization.
• Evaluation using MASBENCH shows robust performance, particularly in parallel and adversarial scenarios, highlighting effective structured delegation.

MAS-Orchestra denotes a line of research on orchestration in multi-agent systems, with its dominant current usage referring to a training-time framework for LLM-based multi-agent reasoning that formulates orchestration as a function-calling reinforcement learning problem with holistic orchestration, generating an entire MAS at once (Ke et al., 21 Jan 2026). In a broader and earlier sense, the term can also denote any “multi-agent system for orchestras,” including category-theoretic descriptions in which musicians, conductor, and listener are modeled through gestures, morphisms, and limit/colimit constructions (Mannone, 2019).

1. Terminological scope and conceptual background

In the broadest sense, MAS-Orchestra can refer to a mathematically precise multi-agent description of an orchestra. In "Introduction to Gestural Similarity in Music. An Application of Category Theory to the Orchestra" (Mannone, 2019), a gesture is defined as a mapping from a directed graph into a space of continuous curves, and orchestral performance is organized through categorical relations among instrumental gestures, conductor gestures, and listener perception. Within that construction, the orchestra is represented as a category $\mathcal{D}$, the conductor as $\mathrm{colim}(\mathcal{D})$, and the listener as $\lim(\mathcal{D})$, yielding a formally explicit conductor–orchestra–listener triad (Mannone, 2019).

The contemporary use of MAS-Orchestra is centered on LLM orchestration. "MAS-Orchestra: Understanding and Improving Multi-Agent Reasoning Through Holistic Orchestration and Controlled Benchmarks" presents MAS-Orchestra as a training-time framework for automatic MAS design and as an empirical program for understanding when multi-agent reasoning is beneficial (Ke et al., 21 Jan 2026). Later work classifies MAS-Orchestra as a training-time MAS method alongside MAS², ScoreFlow, and ToolOrchestra, contrasting it with inference-time systems such as EvoAgent, AOrchestra, and AFlow (Lin et al., 17 Jun 2026). This makes MAS-Orchestra both a specific framework and a reference point in the broader taxonomy of LLM-based orchestration.

2. Holistic orchestration as a function-calling RL problem

The central design move in MAS-Orchestra is to abstract complex, goal-oriented sub-agents as callable functions. The orchestrator does not synthesize their internal code; instead, it emits a global orchestration specification that chooses sub-agents, configures them, and defines their connectivity. If $x$ is the input, $m$ is the Degree of MAS (DoM), and $a$ is the orchestration, then a deterministic parser $f$ executes the orchestration and produces the answer $\hat{y}$ (Ke et al., 21 Jan 2026).

The training objective is defined directly over this orchestration policy:

$$\max_{\theta} \;\mathbb{E}_{(x,y) \sim \mathcal{D}}\;\mathbb{E}_{a \sim \pi_\theta(\cdot \mid x, m)} \left[ R\big(x, y, f(x,a)\big) \right],$$

with binary reward

$$R(x, y, \hat{y}) = \begin{cases} 1, & \hat{y} = y, \ 0, & \text{otherwise}. \end{cases}$$

This setup removes intermediate orchestration steps from the RL horizon and makes the policy operate at the level of complete system design rather than code-level edits (Ke et al., 21 Jan 2026).

DoM controls orchestration capacity. In Low DoM, the orchestrator may use at most one sub-agent and can also choose to solve the task directly. In High DoM, there is no constraint on the number of sub-agents, and the orchestrator can build arbitrary multi-agent graphs subject to the schema (Ke et al., 21 Jan 2026). The reported sub-agent library includes CoTAgent, SCAgent, DebateAgent, ReflexionAgent, and SearchAgent, each exposed through visible signatures while hiding internal prompting or tool-use logic (Ke et al., 21 Jan 2026).

Optimization uses GRPO. For each input, the policy samples a group of candidate orchestrations, executes them, obtains group-normalized rewards, and applies the clipped GRPO objective with KL regularization to a reference policy (Ke et al., 21 Jan 2026). The practical significance of this choice is that orchestration is optimized as a one-shot structural decision rather than as a long sequence of micro-actions.

3. MASBENCH and the controlled study of MAS utility

MAS-Orchestra is paired with MASBENCH, a controlled benchmark designed to determine when multi-agent systems help and when they do not. MASBENCH characterizes tasks along five axes: Depth, Horizon, Breadth, Parallel, and Robustness (Ke et al., 21 Jan 2026).

Axis	Definition
Depth	Length of the longest dependency chain containing the answer
Horizon	Number of intermediate sub-tasks whose answers must be carried forward
Breadth	Maximum in-degree, i.e., maximum number of dependencies of a sub-task
Parallel	Number of independent sub-task components in the graph
Robustness	Number of sub-tasks with attacks

The benchmark is built from controlled task graphs and corresponding natural-language realizations, with train/test splits reported separately for each axis: Depth $\mathrm{colim}(\mathcal{D})$0, Horizon $\mathrm{colim}(\mathcal{D})$1, Breadth $\mathrm{colim}(\mathcal{D})$2, Parallel $\mathrm{colim}(\mathcal{D})$3, and Robustness $\mathrm{colim}(\mathcal{D})$4 (Ke et al., 21 Jan 2026). This construction allows the framework to isolate structural properties that ordinary benchmark suites entangle.

A central result is that MAS gains are not universal. The analysis states that MAS gains depend critically on task structure, verification protocols, and the capabilities of both orchestrator and sub-agents, rather than holding universally (Ke et al., 21 Jan 2026). When sub-agents are relatively weak, MAS outperforms SAS on Horizon, Breadth, and Parallel, while Depth shows little benefit and can underperform SAS. As sub-agents become stronger, MAS gains diminish across Depth, Horizon, Breadth, and Parallel. Robustness is the persistent exception: under adversarial data poisoning, SAS accuracy collapses, whereas MAS retains substantial accuracy because multiple sub-agents and a final moderator can detect and override misleading information (Ke et al., 21 Jan 2026).

Another notable result concerns orchestrator type. Instruction-tuned LLM orchestrators outperform stronger reasoning LLM orchestrators, because the latter tend to produce trivial MAS, often with only one sub-agent and a tendency to solve the problem themselves instead of delegating (Ke et al., 21 Jan 2026). This directly challenges the common assumption that a stronger end-to-end reasoner is automatically a better orchestrator.

4. Benchmark performance and learned orchestration patterns

MAS-Orchestra is evaluated on AIME24, AIME25, HotpotQA, BrowseComp+, and GPQA. The reported Avg@8 results are $\mathrm{colim}(\mathcal{D})$5 on AIME24, $\mathrm{colim}(\mathcal{D})$6 on AIME25, $\mathrm{colim}(\mathcal{D})$7 on HotpotQA, $\mathrm{colim}(\mathcal{D})$8 on BrowseComp+, and $\mathrm{colim}(\mathcal{D})$9 on GPQA (Ke et al., 21 Jan 2026). The listed standalone baselines include CoTAgent, SCAgent, DebateAgent, ReflexionAgent, and SearchAgent, with representative results such as DebateAgent $\lim(\mathcal{D})$0 on AIME24, SearchAgent $\lim(\mathcal{D})$1 on HotpotQA, and SearchAgent $\lim(\mathcal{D})$2 on BrowseComp+ (Ke et al., 21 Jan 2026). Among inference-time orchestration baselines, AFlow reports $\lim(\mathcal{D})$3 on AIME24 and $\lim(\mathcal{D})$4 on GPQA, while MaAS reports $\lim(\mathcal{D})$5 on AIME24 and $\lim(\mathcal{D})$6 on GPQA; among training-time baselines, MAS-GPT reports $\lim(\mathcal{D})$7 on AIME24 and $\lim(\mathcal{D})$8 on GPQA, and ToolOrchestra reports $\lim(\mathcal{D})$9 on AIME24 and $x$0 on GPQA (Ke et al., 21 Jan 2026).

The learned orchestration policies are structurally informative. On AIME24 under Low DoM, after approximately 20 RL steps the orchestrator almost always delegates the entire task to a single sub-agent, primarily selecting ReflexionAgent and DebateAgent (Ke et al., 21 Jan 2026). This suggests that for sequential mathematical reasoning, the best learned policy is often a router over strong single-agent workflows rather than a complex multi-agent topology.

On BrowseComp+ under High DoM, the orchestrator typically uses 3–4 sub-agents, mostly SearchAgents in parallel followed by an aggregation CoT agent (Ke et al., 21 Jan 2026). HotpotQA shows a related but less elaborate pattern: a single SearchAgent is often sufficient, but the orchestrator can still compose multi-step search with summarization when the question requires it (Ke et al., 21 Jan 2026). These learned patterns closely match MASBENCH’s structural findings: parallel search and aggregation benefit from richer MAS, whereas strongly sequential tasks do not.

5. Relation to adjacent orchestration frameworks

Subsequent work situates MAS-Orchestra within a broader family of orchestration methods. Skill-MAS explicitly contrasts training-time MAS, including MAS-Orchestra, with inference-time MAS and proposes a third path in which orchestration knowledge is externalized as a Meta-Skill rather than internalized in model weights (Lin et al., 17 Jun 2026). In that comparison, MAS-Orchestra is described as a system in which the orchestrator is trained via GRPO on function-calling sequences, gaining experience retention through parameters but inheriting the usual training-time scaling constraints (Lin et al., 17 Jun 2026).

AOrchestra represents a different design point. It models any agent as a tuple $x$1, where $x$2 is Instruction, $x$3 is Context, $x$4 is Tools, and $x$5 is Model, and lets a central orchestrator generate sub-agents dynamically through Delegate(\Phi) and Finish(y) actions (Ruan et al., 3 Feb 2026). Unlike MAS-Orchestra’s training-time holistic RL, AOrchestra emphasizes inference-time automatic agent creation and framework-agnostic delegation.

Orchestration Reward Modeling (OrchRM) is directly built on MAS-Orchestra logs. It trains a reward model $x$6 that scores orchestration quality from the query $x$7 and orchestration $x$8 alone, using self-supervised win–lose pairs mined from specialized-over-base and correct-over-incorrect orchestration data (Tsang et al., 11 Jun 2026). OrchRM is reported to improve training efficiency by up to $x$9 in token usage while improving MAS test-time scaling performance by up to $m$0 in accuracy (Tsang et al., 11 Jun 2026). This reframes MAS-Orchestra from a final-answer-reward system into a source of orchestration-level supervision.

Specification-driven work such as MAS-Lab addresses a different layer of the stack. MAS-Lab separates declarative MAS specification, runtime execution, and lab-style validation, aiming to transform MAS from collections of scripts into engineered distributed systems (Augé et al., 29 Jun 2026). In that framing, MAS-Orchestra solves orchestration policy learning, whereas MAS-Lab targets lifecycle continuity, governance, and validation.

6. Limitations, misconceptions, and open directions

A persistent misconception is that MAS universally dominates SAS. The controlled results do not support that claim. MAS-Orchestra shows that MAS gains are concentrated in particular structural regimes, especially Robustness and some Parallel or aggregation-heavy tasks, while strongly sequential tasks such as high-Depth math can favor simpler delegation or even a strong single-agent workflow (Ke et al., 21 Jan 2026).

A second misconception is that a stronger reasoning LLM automatically makes a better orchestrator. The reported experiments show the opposite pattern: stronger reasoning LLM orchestrators often generate trivial MAS with only one sub-agent and a tendency to solve the problem themselves, while instruction-tuned LLM orchestrators produce richer delegation structures and better orchestration behavior (Ke et al., 21 Jan 2026).

The framework also has operational limits. Higher sub-agent reasoning effort often hits context limits, and MAS does not automatically solve long-context failures (Ke et al., 21 Jan 2026). Robustness gains do not emerge automatically from mixed structural training; Robustness requires explicit adversarial-aware training (Ke et al., 21 Jan 2026). More broadly, later reward-modeling work identifies the cost of sub-agent rollouts as the main bottleneck in both orchestrator training and MAS test-time scaling, motivating orchestration-level reward learning as a scalable alternative (Tsang et al., 11 Jun 2026).

Viewed across its broader lineage, MAS-Orchestra names both a concrete LLM orchestration framework and a more general attempt to formalize ensembles as structured multi-agent systems. In the categorical musical setting, this means expressing conductor, orchestra, and listener through gestures, morphisms, limits, and colimits (Mannone, 2019). In the LLM setting, it means learning how to compose specialized agents holistically, and learning when such composition is worthwhile at all (Ke et al., 21 Jan 2026).