Agentic Tuner: Multi-Agent Optimization

• Agentic Tuner is a multi-agent system that automatically optimizes complex parameters by decomposing tasks into modular, role-specific subtasks.
• It employs formal optimization strategies like actor–critic loops and resource-constrained tuning to achieve significant performance improvements in domains such as feedback control and GPU kernel configuration.
• The system ensures interpretability and sample efficiency through standardized inter-agent communication, surrogate predictors, and adaptive online re-tuning.

An agentic tuner is a multi-agent, learning-based system for automatic optimization, configuration, or control of complex parameters in settings where hand-tuning is challenging, expensive, or suboptimal. Agentic tuners coordinate specialized agents—often leveraging LLMs—to decompose the tuning process into structured, modular subtasks, enabling robust, sample-efficient, and interpretable optimization across domains such as feedback control, GPU kernel configuration, reinforcement learning policy selection, and agentic workflow design (Narimani et al., 23 Jun 2025, Qu et al., 19 Jan 2026, Trirat et al., 26 May 2025, O'Callaghan et al., 2021).

1. Architectural Principles of Agentic Tuners

Agentic tuner frameworks follow a modular, multi-agent architecture in which each agent is assigned a well-defined role in the optimization loop. For example, in feedback control design, AgenticControl deploys six specialized LLM agents (LLMSelector, LLMScenarist, LLMActor, LLMCritic, LLMTerminator, LLMJuror), each communicating via strictly-typed JSON schemas. This division enables separation of concerns: controller selection, scenario formulation, parameter search, evaluation, decision logic, and ambiguity resolution are all handled by dedicated agent modules (Narimani et al., 23 Jun 2025).

A similar four-agent closed-loop system governs agentic kernel tuning, with explicit Planning, Generation (semantic and template levels), Tuning, and Testing Agents. The common architecture emphasizes agent specialization, closed-loop iteration, and standardized inter-agent communication (JSON messages or encoding objects), facilitating interpretability and reproducibility (Qu et al., 19 Jan 2026).

2. Mathematical Formulation and Optimization Strategies

Agentic tuners employ formal optimization objectives tailored to their domain. In feedback controller tuning, the search minimizes a composite cost,

$$J(\theta) = \alpha\,\mathrm{MSE}(\theta) + \beta\,T_s(\theta) + \gamma\,M_p(\theta) + \delta\int_0^T|u(t;\theta)|dt$$

subject to closed-loop stability, where $\mathrm{MSE}$ is mean squared error, $T_s$ is settling time, $M_p$ is percent overshoot, and $u(t;\theta)$ the controller-generated signal (Narimani et al., 23 Jun 2025).

Agentic tuners often structure the optimization as an actor–critic loop: an “Actor” agent proposes parameters, while a “Critic” agent simulates/evaluates and signals “EXPLORE” or “EXPLOIT,” with an adaptive schedule shifting from broad exploration to fine local search. Critic feedback drives both acceptance and the degree of future exploration. Decision/Termination agents inspect improvement trends, invoking auxiliary Juror agents if ambiguity or stagnation arises, e.g., recommending refined parameter ranges through RECONSIDER_RANGE actions.

In GPU kernel tuning, search-based optimization is constrained by explicit resource models: $$\theta^* = \arg\min_{\theta: u(\theta;h)\preceq r(h)} T_h(S_\theta, x)$$ where $u(\theta;h)$ denotes parameterized resource usage and $r(h)$ the hardware’s budget vector. Feasible configurations are evaluated on-device, ranked, and top candidates undergo further validation (Qu et al., 19 Jan 2026).

3. Agentic Tuning in Specialized Domains

Feedback Control

AgenticControl demonstrates agentic tuning for feedback controllers (e.g., PID, Full State Feedback) on a range of nonlinear plant systems. The process is robustified by scenario escalation: after nominal operation is achieved, the agentic tuner introduces noise, actuator disturbances, and parametric uncertainties, re-engaging the inner actor–critic loop to ensure solution robustness. A concrete example is PID parameter search, where early iterations sweep parameter boundaries, gradually focusing near historically best gains and adaptively shrinking the range as improvements plateau (Narimani et al., 23 Jun 2025).

GPU Kernel Tuning

Agentic kernel tuning introduces a two-stage pipeline: semantic refactoring followed by template-based parameter optimization. The system first transforms source kernels into parameterizable templates (e.g., tuning block and grid dimensions, unroll factors), then conducts constrained search over feasible parameter space, integrating hardware feedback and performance profiling. The architecture supports backend-agnostic extensions to OpenCL and HIP (Qu et al., 19 Jan 2026).

Multi-Objective RL and Tunable Agents

Agentic-tuner methodology generalizes to multi-objective deep RL by learning a policy family indexed by a preference vector $w \in \Delta^k$. A single function approximator $Q(s, a; w)$ is trained to generalize over $w$, enabling real-time runtime tuning by interpolating agent behaviors (e.g., from cooperative to competitive) via $w$ selection without retraining (O'Callaghan et al., 2021).

Agentic Workflow Optimization

Workflow-level agentic tuners such as Agentic Predictor combine multi-view encoding (graph, code, prompt) with surrogate performance prediction, enabling the ranking and selection of optimal LLM-based agentic workflows before expensive actual execution. Label efficiency is achieved via cross-domain unsupervised pretraining, and low-latency predictors allow efficient search across large workflow spaces (Trirat et al., 26 May 2025).

4. Quantitative Benchmarks and Empirical Gains

Agentic tuners consistently outperform classical tuning and single-agent methods. In control system tuning, AgenticControl achieved an average 55% error reduction in PID tracking compared to MATLAB PIDTuner under nominal and noisy conditions. On the inverted pendulum benchmark, mean squared error was reduced from 0.4566 (PIDTuner) to 0.2694. Comparative LLM study found DeepSeek-V3 achieved cost targets in 34 iterations for FSF control, substantially faster than LLaMA-4-17B and GPT-4o mini (Narimani et al., 23 Jun 2025).

In GPU kernel tuning, the two-stage agentic tuner secured 3.55× speedup (Kernel-1), outperforming pure agent-based rewriting and yielding stable, near-optimal configurations across diverse problem shapes (Qu et al., 19 Jan 2026).

Agentic Predictor’s workflow tuner delivered 84.38% overall prediction accuracy and improved actual workflow performance to 74.43% on AFlow/FLORA-Bench, compared to 62.56% for random search and 71.00% for the best alternative GNN baseline (Trirat et al., 26 May 2025). Ablation studies demonstrate the necessity of multi-view encoding and label-efficient pretraining.

Key results across domains are summarized below:

Domain	Baseline	Agentic Tuner	Improvement
PID Tuning (MSE)	0.4566	0.2694	41% reduction
Workflow Prediction (%)	71.0	74.43	+3.4 points
CUDA Kernel Speedup	2.89×	3.55×	+23%

5. Extensions, Generalization, and Limitations

Agentic tuner techniques generalize to a wide spectrum of domains—any setting where system behavior is governed by high-dimensional, interactively optimized parameters. Kernel tuning architecture is backend-agnostic; agentic RL tuners are compatible with DQN, actor–critic, PPO, SAC, and can be adapted to non-linear scalarization or opponent-modelling. In control, extension to model predictive control or online gain-scheduling is plausible; recurrent agents could allow persistent online adaptation.

Limitations arise from simulation fidelity (control: unmodeled dynamics may break real-world performance), parameter space dimensionality (combinatorial explosion in kernel templates or MPC), and lack of formal optimality/stability certificates (inferred exclusively from empirical metrics). In kernel tuning, success depends on the recognizability of patterns for templateization and accuracy of profiling feedback. Tuning quality may degrade at parameter regimes out of the sampled training/optimization hull (Narimani et al., 23 Jun 2025, Qu et al., 19 Jan 2026, O'Callaghan et al., 2021).

6. Interpretability, Sample Efficiency, and Real-Time Adaptation

A central advantage of agentic tuners is interpretability: parameter search, decision logic, and agent roles are explicit, all communication is schema-guided, and high-level optimization choices map directly back to controllable settings (e.g., kernel parameters, controller gains). In-context learning and reusable multi-agent modules enable rapid online adaptation: AgenticControl, for example, incorporates outcomes of the 10 most recent simulations in each prompt, facilitating immediate re-tuning to environmental changes—without retraining or offline fine-tuning (Narimani et al., 23 Jun 2025).

Sample efficiency is achieved by surrogate predictors, adaptive search, and label-efficient pretraining. In agentic workflow tuning, performance predictors reduce the required number of “expensive” workflow executions by an order of magnitude (Trirat et al., 26 May 2025).

Agentic tuner systems thus establish a systematic, generalizable, and empirically validated framework for automated parameter optimization, bridging expert-driven design with scalable, data-driven learning across contemporary AI and HPC domains.