Symbolic Constraint Agents: Mining Formal Tool-Call Constraints to Prevent Agent Errors

Key points

• Symbolic constraints (programmatic, deterministic checks for tool calls) enable deterministic verification of LLM agents' tool calls, but existing approaches primarily enforce predefined safety policies rather than tool dependencies in complex tool environments.1
• We explore whether symbolic constraints can be mined from environments and observations beyond predefined policies, and whether they can improve end-to-end task performance.1
• On AppWorld, refinement raises error-detection F1 from 37.0 to 58.3 on GPT-5.5 traces and from 39.3 to 61.9 on Qwen3.5-27B traces.1
• These findings show that automatically acquired symbolic constraints can improve tool-calling reliability and task completion, particularly when failures are costly or difficult to diagnose.1

Sources (1)

• [1]Symbolic Constraint Agents: Mining Formal Tool-Call Constraints to Prevent Agent Errors
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 18, 09:49 PM

  We present Symbolic Constraint Agents (SCA), a framework that mines symbolic constraints from API documentation, refines them against execution traces, and checks the agent's tool calls with an SMT solver at runtime.

  Symbolic constraints (programmatic, deterministic checks for tool calls) enable deterministic verification of LLM agents' tool calls, but existing approaches primarily enforce predefined safety policies rather than tool dependencies in complex tool environments.

Extractive summary: sentences quoted from the sources.