Agentic Supernet: Adaptive Multi-Agent Systems

• Agentic Supernet is a probabilistic, dynamic system that represents multi-agent architectures as a continuous supernet over operator selections.
• It employs a Mixture-of-Experts sampling and early-exit strategy to optimize resource allocation, achieving up to 45% cost reduction and improved task accuracy.
• The framework supports cross-model transfer and scalable deployment in real-world applications such as education, code generation, and business intelligence.

An agentic supernet is a probabilistic, dynamically optimized distribution over multi-agent system architectures rather than a rigid, monolithic agentic workflow. By representing the space of possible agentic architectures as a continuous, parameterized supernet, an agentic supernet offers query-adaptive resource allocation, cost-performance trade-off optimization, and superior generalizability across tasks and models, in contrast to traditional one-size-fits-all multi-agent systems.

1. Mathematical Formulation and Conceptualization

The agentic supernet is mathematically formalized as $A = \{T, O\}$, where $T$ denotes a set of parameterized probability distributions over layers (or stages) of an agentic workflow, and $O$ is the library of available operators (e.g., tool-using modules, LLM prompts, external API invocations). Each layer $l$ has a probability $T_l(O)$ over operator inclusion, yielding a flexible, cascaded architecture generation process.

Unlike static pipelines, for every query $q$, the system samples a workflow $G$ by iteratively drawing operators from the supernet's conditional distributions:

$$Q_{\phi}(G|q, T, O) = \prod_{l} T_{l}(V_{l} \mid G_{1:l-1}),$$

where $Q_{\phi}$ is a controller parameterized by $\phi$, and $V_l$ are the operator variants considered at layer $l$. To reduce unnecessary computation on simple queries, an early-exit operator $O_\mathrm{exit}$ is introduced. When sampled, this halts the workflow and outputs an answer immediately.

The agentic supernet thus specifies a stochastic policy over multi-agent system architectures, enabling sample-efficient, task- and budget-aware multi-agent reasoning.

2. The MaAS Framework: Optimization and Execution

The Multi-agent Architecture Search framework (MaAS) operationalizes the agentic supernet by automating three core stages:

1. Architecture Sampling: For each incoming query, the controller $Q_{\phi}$ samples an agentic workflow $G$ from the supernet.
2. Workflow Execution: The sampled architecture $G$ is executed on the given query, producing result $e(a|G)$.
3. Optimization: The agentic supernet is updated to maximize utility and minimize cost using the objective

$$\max \mathbb{E}_{(q,a)}\left[ U(G; q, a) - \lambda C(G; q) \right]$$

where $U$ measures downstream utility (e.g., answer accuracy), $C$ quantifies inference costs (e.g., LLM calls, token budget), and $\lambda$ is a user-tunable trade-off parameter.

Formally, the conditional output probability is integrated over all architectures:

$$p(a\mid q,T,O) = \int e(a|G)\, Q_{\phi}(G|q,T,O)\, dG.$$

MaAS employs a Mixture-of-Expert (MoE) style operator selection: at each layer, operator activations are computed via embeddings, and sampling is based on query-conditioned activation scores. This allows the system to allocate computation proportional to perceived query complexity.

3. Resource Allocation and Cost Control

A primary advantage of the agentic supernet is dynamic, per-query resource allocation:

• For simple queries, the supernet can generate shallow or even single-step architectures, minimizing LLM/tool invocations and token usage.
• For complex queries, multi-layer, multi-operator workflows can be sampled, providing deep reasoning at higher, but justified, resource cost.
• The early-exit operator $O_\mathrm{exit}$ allows the system to terminate reasoning when sufficient confidence is achieved.

The cost-performance objective integrated in MaAS makes explicit the Pareto frontier between answer quality and computation, adjustable via $\lambda$. This is essential for scalable, budget-constrained deployments.

4. Empirical Evaluation and Performance Metrics

The agentic supernet and MaAS were benchmarked across six datasets spanning math reasoning (GSM8K, MATH, MultiArith), code generation (HumanEval, MBPP), and tool-augmented scenarios (GAIA). Notable results include:

• Inference Cost: MaAS requires only $6\sim45\%$ of the computational cost (API/tokens, wall-clock time) compared to static multi-agent baselines.
• Performance: Outperforms previous handcrafted or automated multi-agent systems by $0.54\%\sim11.82\%$ in task-specific accuracy and pass@1 metrics.
• Efficiency: On representative tasks such as MATH, MaAS achieves lowest total token usage and shortest inference time while attaining highest accuracy.

These metrics demonstrate that agentic supernets deliver both computational efficiency and solution quality, a property unattainable by monolithic or statically optimized agentic systems.

5. Transferability and Scalability

A key feature of the agentic supernet is robust transferability:

• Cross-LLM Backbones: The supernet optimized with one LLM backbone (e.g., gpt-4o-mini) retains its performance gains when deployed with alternative LLMs (e.g., Qwen-2.5-72b, llama-3.1-70b) with no retraining.
• Cross-Domain Generalization: A supernet trained on one benchmark can perform well on others (e.g., optimized on MATH and applied to GSM8K).
• Scalability: The probabilistic, continuous architecture space represented by the agentic supernet generalizes effectively to new subtasks and operator libraries.

This suggests high scalability in settings where user queries and task domains evolve rapidly.

6. Real-World Applications

The adaptive and resource-aware properties of agentic supernets enable several practical deployment modes:

• Personalized educational assistants: Sampling task-specific multi-agent workflows for mathematical education, adjusting resource expenditure based on student query complexity.
• Industrial code generation: Cost-adaptive code assistant that routes simpler snippets to lightweight workflows and reserves deep, iterative reasoning for complex programming tasks.
• Integrated multi-agent problems: On-the-fly combination of tools—CoT, ReAct, Debate, external APIs—by sampled architectures tuned to user needs, domain, and budget.
• Business intelligence and support: Dynamically scales reasoning steps and cost in response to business user query complexity.

These patterns allow for efficient, scalable deployment in heterogeneous, real-world environments.

7. Technical and Practical Considerations

From an implementation perspective, agentic supernets and MaAS introduce considerations including:

• Controller Complexity: The architecture-sampling controller must efficiently sequence operator selection, support early-exit, and represent multi-layer dependencies.
• Operator Library Expansion: The supernet is easily extensible—new operators can be added and become accessible to the search space without retraining the entire system.
• Computational Overhead: The sampling process and MoE activations introduce small overheads, but these are offset by the substantial overall inference cost reduction.
• Deployment Strategy: The model can be tuned for aggressive cost-minimization or high-accuracy regimes by adjusting the trade-off parameter $\lambda$.

Potential limitations include architectural search complexity in extremely large operator spaces and sensitivity to initial operator library composition.

---

In summary, the agentic supernet paradigm, instantiated via MaAS, replaces static agentic system design with a probabilistically defined, query-adaptive workflow distribution. This enables dynamic resource allocation, cost-performance trade-off optimization, robust cross-task and cross-model transfer, and superior scalability, as validated by empirical benchmarks in mathematical reasoning, code generation, and tool-based multi-agent scenarios.