RAC Introduces Relational Contracts for Constraint-Governed Engineering Agents

Summary

Engineering agents often have to decide what an artifact should satisfy and also derive exact geometry, timing, or other consequences, which can cause failures in tightly constrained tasks. The paper introduces RAC, a harness-agnostic contract for the action and feedback layer of engineering agents. RAC defines authoritative observations, semantic actions, tool-based realization, independent verification, and repair evidence. Under the contract, the agent chooses the relations an artifact should satisfy, while specialized tools derive concrete consequences and verification checks the result. On PCB placement, routing, and repair tasks, RAC produced the highest clean-pass rate for all nine evaluated models, averaging 90.9% versus 74.0% for the strongest requirements-only delegation baseline. In job-shop scheduling, direct relational actions achieved a 96.3% quality pass rate, compared with 59.6% for code-mediated delivery using the same scheduling backend. The study also evaluates 141 embedded-code tasks across four microcontrollers. Its results attribute the reliability improvement to the action contract: preserving semantic control while delegating mechanically derivable outputs can improve the dependability of constrained engineering artifacts.