El Agente Forjador: A Task-Driven Agent Tool Generation Framework for Quantum Simulation

The El Agente Forjador framework enables general-purpose coding agents to autonomously forge, validate, and reuse computational tools through a four-stage workflow. It achieves dual optimization of accuracy and cost in quantum chemistry and quantum dynamics tasks, breaks the bottleneck of traditional static tool sets, and opens up a new path for AI for Science.

Artificial intelligence accelerates scientific tasks, but existing agents rely on static manually curated tool sets, which have three major limitations: difficulty adapting to new domain requirements, inability to keep up with library version updates, and lack of tool reuse mechanisms. This dilemma is particularly prominent in the quantum science field due to complex computing and professional knowledge requirements, limiting the application potential of AI.

The framework adopts a multi-agent architecture, where general-purpose coding agents manage the tool lifecycle through four stages:

1. Tool Analysis: Identify the computational requirements, algorithm types, and library dependencies of the task;
2. Tool Generation: Write reusable components with interface design, error handling, and documentation;
3. Task Execution: Use the generated tools to solve actual tasks, validate functions, and collect performance data;
4. Iterative Evaluation: Evaluate results, iteratively optimize tools, and include them in the reuse set after validation.

Three modes were compared across 24 quantum chemistry/dynamics tasks:

• Zero-shot generation: Flexible but high cost;
• Curriculum-based reuse: Prioritize reuse of existing tools;
• Baseline: Direct generation without tools using coding agents. Results show that the framework consistently outperforms the baseline. Reusing tools allows weak agents to reduce API call costs and improve problem-solving quality, and tools can serve as knowledge carriers to transfer between agents of different capabilities.

Case studies found that tools from different domains can be combined to solve hybrid tasks (e.g., quantum chemistry combining molecular structure analysis and quantum state evolution tools). Traditional static tool sets require manual predefined combinations, while the framework can dynamically identify needs and automatically coordinate tools, marking a key step toward agent generality.

The framework needs to address three key technical challenges:

• Tool Representation: Design standardized interfaces;
• Tool Retrieval: Fast matching via metadata (functionality, input/output, applicable scenarios) + semantic retrieval;
• Tool Validation: Ensure reliability through unit testing, property checking, and result consistency verification.

The framework promotes a paradigm shift: agent capabilities are defined by tasks rather than explicit engineering. The speed of scientific research iteration is accelerated and the threshold is lowered; researchers can directly describe their needs for agents to generate tools. The "demand as tool" model is expected to accelerate scientific discovery.

Limitations: The quality of tool generation depends on the programming ability of the underlying agent; complex algorithms require human intervention; long-term maintenance and version management of tool sets need further research. Future directions: Efficient tool learning strategies, tool dependency management and conflict resolution, and expansion to more scientific fields to verify generality.