TopOptAgents: Agentic Optimization Frameworks

• TopOptAgents are a family of agentic optimization frameworks that assign distinct cognitive and computational roles to specialized agents.
• They coordinate tasks like formulation, evaluation, execution, and critique through iterative self-refinement and structured memory systems.
• Applied in hyperparameter and topology optimization, TopOptAgents improve success rates and offer a novel mapping of expert heuristics into automated workflows.

Searching arXiv for the cited papers and related "TopOptAgents" usage. arXiv search: "TopOptAgents". TopOptAgents denotes a family of agentic optimization frameworks in which specialized agents decompose an optimization workflow into separable cognitive and computational roles such as formulation, proposal, evaluation, execution, critique, and strategic control. In the recent arXiv literature, the term appears in two closely related senses. In "OptiMindTune" it refers to specialized “Top Optimization Agents” for hyperparameter optimization (HPO), organized around recommendation, evaluation, and decision-making (Madiraju et al., 25 May 2025). In "Self-Refining Topology Optimization via an LLM-Based Multi-Agent Framework" it is the explicit name of a multi-agent system for automating topology optimization through iterative self-refinement (Park et al., 22 May 2026). This suggests that TopOptAgents is evolving from a role label for optimization-specialized agents into a broader design pattern for end-to-end, LLM-mediated optimization systems.

1. Terminology and historical usage

The earliest usage in the present corpus is tied to HPO. OptiMindTune frames HPO as a cooperative multi-agent problem in which several specialized “Top Optimization Agents” share information, reason with Gemini, and coordinate to explore and exploit a difficult search space more intelligently than a single algorithmic optimizer (Madiraju et al., 25 May 2025). In that formulation, “TopOptAgents” is not a standalone platform name but a role-based abstraction: proposal, evaluation, and strategic selection are separated into distinct agents.

By 2026, the term acquires a narrower and more formal meaning in computational mechanics. TopOptAgents becomes the name of a topology-optimization framework that automates not only numerical execution but also expert decision-making during problem formulation, validation, code generation, execution, and post-convergence quality assessment (Park et al., 22 May 2026). In that usage, the “TopOpt” component refers to topology optimization rather than generic optimization.

A distinct but related terminological neighbor appears in TopoDIM, where “topology optimization” refers to optimizing communication topology in LLM-based multi-agent systems. There the system is modeled as a directed heterogeneous graph whose edges encode interaction modes such as conditioned, feedback, and debate; this is not topology optimization in the structural-design sense, but topology optimization of agent communication itself (Sun et al., 15 Jan 2026). The overlap in terminology is substantive rather than accidental: both lines of work treat optimization as a problem of coordinating heterogeneous roles under constrained compute and uncertain feedback.

2. Architectural pattern

Across the literature, TopOptAgents-style systems share three recurrent architectural ideas: specialization, explicit coordination, and structured memory. Specialization assigns different subtasks to distinct agents; coordination determines how proposals, executions, and critiques are routed; structured memory records prior trials, traces, or artifacts so that later decisions are history-aware rather than stateless. In OptiMindTune, this appears as a centralized message-passing workflow with shared logs. In TopOptAgents for topology optimization, it appears as a structured memory state combined with Pydantic-enforced outputs. In TO-Master, the same logic is realized through typed Model Context Protocol tool calls and artifact-aware dialogue state (Madiraju et al., 25 May 2025, Park et al., 22 May 2026, Lin et al., 2 Jul 2026).

System	Roles	Coordination pattern
OptiMindTune	Recommender, Evaluator, Decision	Centralized message-passing workflow
TopOptAgents	Scientist, Validator, Planner, Coder, Executor, Reviewer, Critic	Three internal self-refinement loops plus a user-driven loop
TO-Master	LLM agent with typed MCP tools	Tool-orchestrated workflow with BC confirmation before solver launch

In OptiMindTune, the Recommender Agent acts as intelligent explorer and knowledge synthesizer, the Evaluator Agent acts as a robust and reproducible evaluation engine, and the Decision Agent acts as strategic coordinator and meta-optimizer. Their coordination is explicit: Recommender proposes configurations, Decision selects one, Evaluator returns cross-validated metrics, and Decision then updates global state and feeds back an exploration or exploitation directive (Madiraju et al., 25 May 2025).

In TopOptAgents for topology optimization, the decomposition is more granular because the workflow contains additional failure modes. Scientist formulates the optimization problem from natural language, Validator performs sanity checks and corrections, Planner decomposes code generation into function-level tasks, Coder writes Python code, Executor runs it as a subprocess, Reviewer diagnoses runtime behavior, and Critic evaluates the final design and routes refinements upstream (Park et al., 22 May 2026). The result is not merely distributed execution but distributed expertise.

TO-Master shows a different architectural choice. It is presented as an LLM-agent framework rather than a multi-agent society, but it instantiates the same principle of separating language-level orchestration from deterministic numerical work. The LLM never solves the PDE or performs numerical optimization directly; it selects tools, constructs typed arguments, checks meshes and boundary conditions, and launches JAX-FEM-based topology optimization only after explicit boundary-condition confirmation (Lin et al., 2 Jul 2026).

3. Formal optimization models and agent mappings

The TopOptAgents literature does not replace standard optimization formulations; rather, it overlays an agentic control layer on top of them. In OptiMindTune, HPO is stated as

$$\lambda^\star = \arg\min_{\lambda \in \Lambda} \mathcal{L}(\lambda),$$

or equivalently as maximization of validation accuracy over a joint hyperparameter space that may span multiple model families. The Recommender Agent proposes candidate configurations $\lambda$, the Evaluator Agent estimates performance by $k$-fold cross-validation,

$$\hat{A}(\lambda) = \frac{1}{k}\sum_{i=1}^k A_i(\lambda),$$

and the Decision Agent operates on the history $\{(\lambda_i,\hat{A}(\lambda_i))\}_{i=1}^t$ to select the next action and decide whether to terminate (Madiraju et al., 25 May 2025).

In topology optimization, the base problem is instead a PDE-constrained design problem. TopOptAgents writes the general form as

$$\min_{\mathbf{x},\mathbf{s}} f(\mathbf{s},\mathbf{x}) \quad \text{s.t.}\quad g_i(\mathbf{s},\mathbf{x}) \le 0,\; h_j(\mathbf{s},\mathbf{x}) = 0,\; \mathbf{x}\in\mathcal{X}_{\text{ad}},$$

with design variables, state variables, objectives, constraints, and admissible sets made explicit. For density-based structural problems, the PDE constraint is given after FEM discretization by

$$\mathbf{R}(\mathbf{u},\boldsymbol{\rho})=\mathbf{K}(\boldsymbol{\rho})\mathbf{u}-\mathbf{F}=\mathbf{0},$$

and sensitivities are computed through an adjoint formulation (Park et al., 22 May 2026).

TO-Master instantiates this same formalism through density-based structural and thermal topology optimization. For structural compliance minimization it uses SIMP interpolation,

$E_e = E_{\min} + (E_{\max}-E_{\min})\rho_e^p,$

a volume fraction constraint, density filtering, optional Heaviside projection, and MMA updates. For thermal topology optimization it replaces elastic constitutive interpolation with conductivity interpolation and minimizes heat-source-weighted temperature. For stress-constrained problems it adopts an aggregated, density-weighted von Mises stress measure with a $p$-norm global stress constraint (Lin et al., 2 Jul 2026).

What the agentic layer contributes is not a new objective class but a new mapping from semantic instructions and intermediate evidence to optimization actions. In HPO, this mapping is an LLM-infused policy over candidate hyperparameters and search direction. In topology optimization, it is a policy over formulation choices, numerical parameters, code repairs, and post-convergence design-quality refinements. A plausible unifying interpretation is that TopOptAgents externalizes expert heuristics that are usually implicit in manual optimization workflows.

4. Realizations in topology optimization

The explicit TopOptAgents framework for topology optimization is organized around iterative self-refinement. The workflow begins with a natural-language user query, which Scientist translates into a structured optimization specification: objective, constraints, PDE, design and state variables, domain, boundary conditions, required numerical techniques, and default parameters when omitted. Validator then checks consistency with the user request, numerical completeness, and boundary-condition sanity. If the issue is local, Validator directly edits the specification; if it is structural, Validator requests reformulation from Scientist. This loop repeats until approval (Park et al., 22 May 2026).

After validation, Planner decomposes the implementation into function-level tasks such as mesh construction, material interpolation, FEM assembly in FEniCS, adjoint sensitivity computation, filtering and projection, optimizer setup via pyOptSparse, and the main optimization loop. Coder synthesizes a self-contained Python script, Executor runs it, and Reviewer inspects stdout, stderr, generated plots, and numerical pathologies. When the program runs to convergence, Critic applies a four-stage rubric: output validity, problem-formulation consistency, convergence, and design quality. If any criterion fails, Critic identifies the responsible upstream stage and triggers another refinement cycle (Park et al., 22 May 2026).

A central feature of this system is that it distinguishes failures by informational level. Validator sees specifications but not runtime data. Reviewer sees runtime errors but not final design quality. Critic sees final designs, convergence plots, and optimizer flags. This division is important because topology optimization often fails in ways that are invisible to ordinary code-generation loops: a program may run, converge, and still encode the wrong objective, the wrong load position, checkerboarding, disconnected members, or excessive gray regions (Park et al., 22 May 2026).

TO-Master realizes a related but more tool-centric form of automation. From natural language plus optional mesh, geometry, or image inputs, the agent selects computational tools, constructs finite-element topology-optimization models, checks meshes and boundary conditions, and launches sensitivity-based optimization with typed solver arguments. The framework supports generated and uploaded meshes, image-to-mesh conversion, 2D and 3D structural compliance minimization, thermal conduction, multiple load cases, stress-constrained optimization, and engineering geometries. Its numerical backend is JAX-FEM with MMA, while tool invocation and artifact management are mediated through Google ADK, MCP, and bohr-agent-sdk (Lin et al., 2 Jul 2026).

TO-Master places unusual emphasis on boundary-condition verification. Its system prompt enforces that mesh and boundary conditions must be visualized before optimization, and that only after user confirmation may the tool set bc_confirmed=True and run topology optimization. This choice reflects a recurrent observation in agentic topology optimization: many failures originate not in the optimizer itself but in translating informal spatial language such as “lower-right corner,” “middle part of the left boundary,” or “four lower bolt holes” into solver-valid geometric selections (Lin et al., 2 Jul 2026).

5. Empirical performance and evaluation

In HPO, OptiMindTune evaluates its TopOptAgents formulation on three UCI datasets: Breast Cancer Wisconsin (Diagnostic), Iris, and Wine. The objective is 5-fold cross-validated accuracy, with Optuna as the baseline under a 10-trial budget. OptiMindTune determines its number of trials adaptively and uses 3–4 trials in the reported runs. It achieves 97.02% on Breast Cancer versus Optuna’s 96.14%, 96.67% on Iris versus Optuna’s 98.00%, and 98.33% on Wine versus Optuna’s 97.78%. It therefore wins on 2 out of 3 datasets despite Optuna being allowed to choose the better of two models per dataset. The same experiments show the expected trade-off between sample efficiency and per-trial overhead: OptiMindTune uses fewer trials, but LLM calls and coordination reduce trials per second and can increase wall-clock time on some datasets (Madiraju et al., 25 May 2025).

For topology optimization, TopOptAgents is benchmarked against a single LLM baseline, OpenAI GPT-5.4 Thinking, on three 2D problems: cantilever compliance minimization, modified MBB beam compliance minimization, and an L-shaped stress problem. Each method is run for 10 independent sessions per problem. The reported success rates are 100% for both methods on cantilever compliance, 100% for TopOptAgents versus 80% for the single LLM on the modified MBB beam, and 80% versus 10% on the L-shaped stress problem. The number of self-refinement iterations also tracks problem difficulty: 0 for cantilever, a small number for the modified MBB case, and up to 12 for the L-shaped stress problem (Park et al., 22 May 2026).

TO-Master evaluates robustness by ablation over prompt structure on six benchmark cases: 2D cantilever, 3D cantilever, 2D bridge, 3D bridge, 2D thermal conduction, and a 3D multi-load tower. With the full prompt configuration—role, tool-usage rules, internal reasoning guidance, and few-shot examples—the success rate is 96.7%. Removing chain-of-thought guidance reduces this to 83.3%; removing few-shot examples reduces it to 73.3%; and removing tool-usage constraints reduces it to 40.0%. The dominant failure modes are passive-domain misinterpretation and boundary/load misinterpretation. This establishes that, for TopOptAgent-style systems, prompt architecture and tool-usage constraints are operational components rather than cosmetic interface choices (Lin et al., 2 Jul 2026).

Adjacent work on agent optimization reinforces this conclusion from a different angle. VeRO shows that agents which optimize other agents benefit materially from a reproducible evaluation harness with versioned agent snapshots, budget-controlled evaluation, and structured execution traces; on its benchmark suite, the best VeRO scaffold configuration improves the average performance from a 0.50 baseline to 0.61, with 0.65 as the best maximum across seeds (Ursekar et al., 25 Feb 2026). While VeRO is not a topology-optimization framework, it provides empirical support for the broader claim that optimization agents require explicit instrumentation, not just a stronger base model.

6. Relations to adjacent research, limitations, and open questions

TopOptAgents sits at the intersection of classical optimization, LLM-based agent orchestration, and multi-agent systems. In HPO, it differentiates itself from grid search, random search, Bayesian optimization, evolutionary algorithms, and Hyperband-style resource allocation by replacing monolithic control logic with specialized agents that can reason about model families, dataset characteristics, and search history in natural language (Madiraju et al., 25 May 2025). In topology optimization, it extends earlier AI-assisted engineering pipelines by making post-convergence quality assessment and targeted re-entry into earlier stages first-class operations (Park et al., 22 May 2026).

Several adjacent frameworks illuminate possible future directions. TopoDIM treats communication topology as a stochastic policy over directed heterogeneous graphs and shows that one-shot, heterogeneous topology generation with conditioned, feedback, and debate edges can reduce total token consumption by 46.41% while improving average performance by 1.50% over state-of-the-art methods (Sun et al., 15 Jan 2026). This suggests that future TopOptAgents need not remain centrally orchestrated; learned communication structures could replace fixed pipelines or hand-authored message routing. AT$^2$PO addresses multi-turn agentic reinforcement learning through a turn-level tree, entropy-guided tree expansion, turn-wise credit assignment, and a turn-based policy objective, with gains of up to 1.84 percentage points in average over the state-of-the-art baseline across seven benchmarks (Zong et al., 8 Jan 2026). A plausible implication is that TopOptAgents could eventually learn coordination policies, not only execute prompt-engineered role assignments.

The present literature also converges on a common set of limitations. OptiMindTune emphasizes dependency on LLM services, opacity, communication overhead, and difficulty with fine-grained continuous optimization relative to dedicated numerical methods (Madiraju et al., 25 May 2025). TopOptAgents for topology optimization notes dependence on pretrained LLM knowledge, a coverage gap for exotic formulations, computational cost from repeated refinement cycles, and continued failures on the hardest stress problem (Park et al., 22 May 2026). TO-Master adds scope limits in physics, reliance on its JAX-FEM-based backend, residual interpretation errors under ambiguous prompts, and the absence of automatic unit conversion (Lin et al., 2 Jul 2026).

Open questions follow directly from these constraints. The topology-optimization paper asks how to formally characterize optimization behavior when the search policy is an LLM rather than an explicit acquisition function, how to design interpretable communication protocols between agents, how to scale to large and time-critical settings where LLM calls are costly, and how to ensure reproducibility under stochastic LLM outputs (Park et al., 22 May 2026). OptiMindTune proposes more advanced LLM integration, reinforcement learning for coordination, cost-aware multi-objective optimization, and transfer or meta-learning across tasks (Madiraju et al., 25 May 2025). Taken together, these results position TopOptAgents less as a settled framework than as a research program: agent specialization, structured self-refinement, typed tool orchestration, and explicit evaluation are already technically grounded, whereas formal guarantees, scalable coordination, and cross-task generalization remain open.