TopoPilot: Reliable Conversational Workflow Automation for Topological Data Analysis and Visualization

Researchers Nathaniel Gorski, Shusen Liu, and Bei Wang have introduced TopoPilot, a new agentic framework designed to solve the critical reliability problem in AI-driven scientific visualization. Current systems powered by large language models (LLMs) often execute invalid operations, introduce subtle errors, or fail to ask for missing information, especially in complex, real-world workflows that go beyond simple benchmarks. TopoPilot tackles this by implementing a reliability-centered, two-agent architecture. An orchestrator agent interprets a user's natural language prompt and translates it into a workflow of atomic backend actions. A separate verifier agent then rigorously evaluates this proposed workflow for structural validity and semantic consistency before any code is executed. This separation of interpretation and verification is key to reducing code-generation errors and enforcing correctness guarantees.

To systematically address failure modes, the team created a taxonomy of potential errors and implemented targeted safeguards for each class. The framework's modular design also isolates components, making it extensible and allowing for the seamless integration of new domain-specific workflows without altering the core system. While the primary use case demonstrated is topological data analysis (TDA)—a complex field for understanding data shape and structure—the architecture is built to generalize across visualization domains. In rigorous evaluations simulating 1,000 multi-turn conversations across 100 prompts, including adversarial and infeasible requests, TopoPilot achieved a remarkable success rate exceeding 99%. This performance starkly contrasts with baseline systems lacking comprehensive guardrails, which succeeded less than 50% of the time. The work represents a significant step toward trustworthy, autonomous AI assistants for technical and scientific work.