Planner-Executor Architecture in Scientific Workflows

• Planner-Executor Architecture is a modular framework that decomposes high-level tasks into hierarchical subtasks executed by specialized agents.
• It enables dynamic workflow adaptation by integrating LLM-driven planning with tailored toolsets for tasks like materials discovery and catalyst design.
• The framework improves efficiency and scalability in complex scientific workflows by automating experiment coordination, data analysis, and performance optimization.

A Planner-Executor Architecture is a modular agentic framework in which a central planning module decomposes a complex, high-level task into a hierarchy of subtasks, and multiple specialized executor modules (agents) carry out each subtask with tailored toolsets and configurations. This architectural pattern eliminates the need for static, manually defined workflows and enhances adaptability, scalability, and efficiency, particularly in domains such as inverse materials design, where complex multi-stage reasoning and experimental validation must be robustly coordinated (Wang et al., 18 Sep 2025).

1. Structural Overview of the Planner-Executor System

In the S1-MatAgent system, the Planner-Executor architecture is instantiated as an LLM-driven multi-agent framework for automated materials discovery:

• Planner Agent: Acts as an intelligent scheduler, receiving a high-level root task (e.g., “Design high-entropy alloy catalysts for HER in alkaline conditions”) and autonomously decomposing it into a hierarchical task network (HTN). The HTN encodes both task hierarchies (compounding and primitive subtasks) and their temporal/causal dependencies.
• Executor Agents: Each primitive subtask in the HTN triggers dynamic instantiation of a dedicated executor agent. Executors are configured with a specialized toolset (such as literature mining modules, code generation tools, or simulation/model evaluation tools) and operate under a system message specifying their precise role and requirements.
• Working Memory: The Planner maintains a working memory, recording the tree structure, dependency relationships, and all intermediate results. This enables aggregation of partial results and systematic updating of the workflow as subtasks are completed.

This approach enables closed-loop, full-cycle automation: from literature extraction and statistical analysis, through composition recommendation and optimization, to experimental validation.

2. Planner Functionality and Workflow Construction

The Planner leverages LLM-based reasoning to translate user requests into structured workflow graphs:

• Task Decomposition: The planner analyzes the root design request, consults the available tool "inventory," and determines whether to solve the task directly or decompose it into compound and primitive subtasks.
• HTN Generation: It generates a hierarchical task network where nodes represent tasks (compound or primitive) and edges encode temporal or logical dependencies. Primitive tasks correspond to actions that can be delegated to executor agents.
• Agent Instantiation and Specialization: For each primitive subtask, the planner creates a new executor agent instance, attaches a specialized system prompt detailing the subtask, and configures the required tools.
• Result Aggregation: As each executor completes its assignment, results are written to the working memory. The planner then synthesizes these results, potentially spawns further subtasks, and ultimately delivers a final integrated solution.

The planning process is highly adaptable: new tasks, tools, or requirements can be dynamically integrated, and the workflow is autonomously restructured in response to input changes or execution feedback.

3. Executor Agent Design and Specialization

Executor agents in S1-MatAgent are dynamically generated, subtask-specific agents closely tied to the Planner’s HTN decomposition:

• Dedicated Toolsets: Each Executor is provisioned with task-relevant tools—for example:
  • Literature extraction agents with chemical formula parsing and regular expressions for text mining.
  • Statistical analysis agents with code generation capabilities for element frequency quantification.
  • Simulation agents with access to DFT, MLIP (machine learning interatomic potentials), and performance prediction frameworks.
• Parallel/Sequential Execution: Executors work in both parallel and sequential modes, depending on HTN dependencies. For example, literature extraction and frequency analysis may proceed in parallel, with their outputs aggregated before composition recommendation.
• Specialized System Messages: System prompts are tailored to maximize agent focus, minimize context bloat, and ensure robust subtask completion.

This dynamic instantiation and modular role segregation enable efficient management of large-scale, heterogeneous workflows and facilitate easy scaling to new scientific domains.

4. Application to High-Entropy Alloy Catalyst Discovery

S1-MatAgent applies this architecture to automate the discovery of catalysts for the hydrogen evolution reaction (HER):

• Literature Mining: An Executor agent parses 1,231 scientific articles to extract chemical formulas, leveraging regex-based extraction tools.
• Statistical Filtering: Statistical Executor agents count element frequencies, narrowing a vast 20-million-candidate search space to a practical subspace by identifying top-10 element candidates.
• Composition Recommendation: Using ScienceOne and other advanced models, a recommendation Executor proposes initial high-entropy alloy (HEA) compositions from the refined search space.
• Performance Optimization: Optimization Executors perform gradient-based refinement of candidate compositions using machine learning interatomic potentials (MLIP), specifically computing the gradient of the predicted activity descriptor with respect to the HEA’s elemental representation.
• Closed-Loop Feedback: Each stage’s results (e.g., updated candidate lists, predicted activities) are fed back to the Planner and stored in the working memory for further planning or experimental validation.

This workflow enabled the agent to design and optimize 13 high-performance HEA catalysts, from which top candidates were experimentally validated: Ni4Co4Cu1Mo3Ru4 achieved an overpotential of 18.6 mV at 10 mA/cm² and maintained 97.5% activity after 500 hours at 500 mA/cm².

5. Optimizing Performance via Gradient-Based Composition Algorithms

A distinctive aspect of S1-MatAgent is the use of MLIP-based gradient optimization to steer alloy composition refinement:

• Gradient Computation: Post-recommendation, the MLIP model (built on architectures such as MACE) provides a computational graph from which the activity descriptor’s gradient with respect to the HEA’s composition is computed:

$$f_{\text{new}} \approx f_{\text{current}} + \nabla f \cdot \Delta x$$

where $f$ is the activity, $\nabla f$ is the gradient, and $\Delta x$ is the vector of composition changes.

• Refinement Strategies:
  • Adjusting Element Ratios: Reduce the proportion of low-contribution elements and increase more positive contributors.
  • Changing Element Types: Substitute all atoms of one element with another higher-contributing candidate.
• Performance Gains: Applying these strategies systematically increased mean predicted HER activity from –21.12 to –15.28, equating to a 27.7% improvement.

6. Scalability, Adaptability, and Impact

The Planner-Executor framework in S1-MatAgent enables:

• Dynamic Workflow Adaptation: Any new material design request can be decomposed into a fresh HTN, with subtasks and Executors dynamically instantiated as needed—including the integration of new tools or techniques without manual workflow reconfiguration.
• Modularity of Executor Agents: Task-specific customization is trivial; as new scientific tools emerge, Executors can be updated or replaced independently and immediately leveraged by the planner.
• Closed-Loop Iteration: The architecture supports iterative refinement and rapid re-planning, bridged by the working memory, which maintains persistent task and result hierarchies across the design cycle.
• Universality and Generalizability: While demonstrated on HEA catalyst discovery, nothing in the architecture precludes extension to battery materials, solid electrolytes, photonic systems, or any scientific design task that can be hierarchically decomposed.

7. Results and Broader Implications

Empirical evaluation on the full-scale alloy discovery task demonstrated:

• Efficiency: S1-MatAgent successfully searched 20 million candidates, recommending a final set of 13 validated high-performance catalysts, autonomously bridging computation and experiment—results: overpotential 18.6 mV at 10 mA/cm², 97.5% activity retention at 500 hours.
• Performance Optimization: The gradient-based MLIP optimization yielded 27.7% improvement in predicted activity.
• Adaptability: The dynamic planner-driven agent system lowered dependency on manual workflow construction and tool configuration, offering inherent scalability and extensibility.

A plausible implication is that such architectures can accelerate scientific innovation by automating the “last mile” of cross-disciplinary workflow composition and optimization, substantially reducing barriers to rapid hypothesis evaluation and experimental feedback in computational and experimental sciences.

---

Summary Table: Major Components of S1-MatAgent Planner-Executor Architecture

Component	Role	Key Method
Planner Agent	Task decomposition, HTN generation, agent provisioning	LLM-based hierarchical reasoning
Executor Agents	Subtask execution with specialized tools	Dynamically configured LLM/tool pipelines
Working Memory	State, results, dependencies recording	Hierarchical tree update and result tracking
Gradient Optimizer	Performance optimization for compositions	MLIP-based gradient descent/heuristics

S1-MatAgent's Planner-Executor architecture demonstrates that integrating dynamic, LLM-powered planning and modular, tool-specific agent instantiation considerably advances the efficiency, scalability, and robustness of automated scientific discovery workflows (Wang et al., 18 Sep 2025).