RESEARCH · RESEARCH · #2089
RAC: Relational Agent Contract for constraint-governed engineering tasks (arXiv:2610.20935v1)
The paper introduces RAC, a harness-agnostic action/feedback contract for engineering agents that separates semantic decisions from mechanically derived realizations by specifying authoritative observations, semantic actions, tool-based realization, independent verification, and repair evidence. Evaluations show RAC improves reliability on constrained engineering tasks—achieving a 90.9% average clean-pass rate on PCB placement, routing, and repair (versus 74.0% for a requirements-only delegation baseline), 96.3% quality pass on job-shop scheduling (versus 59.6% for code-mediated delivery), and gains on 141 embedded-code tasks across four microcontrollers.
KEY POINTS
- The paper introduces RAC, a harness-agnostic action/feedback contract for engineering agents that separates semantic decisions from mechanically derived realizations by specifying authoritative observations, semantic actions, tool-based realization, independent verification, and repair evidence.
- Evaluations show RAC improves reliability on constrained engineering tasks—achieving a 90.9% average clean-pass rate on PCB placement, routing, and repair (versus 74.0% for a requirements-only delegation baseline), 96.3% quality pass on job-shop scheduling (versus 59.6% for code-mediated delivery), and gains on 141 embedded-code tasks across four microcontrollers.
- RAC demonstrates that preserving semantic control while delegating mechanically derivable outputs to tools substantially increases reliability of constraint-governed engineering artifacts, highlighting an actionable design for more capable engineering agents.
WHY IT MATTERS
RAC demonstrates that preserving semantic control while delegating mechanically derivable outputs to tools substantially increases reliability of constraint-governed engineering artifacts, highlighting an actionable design for more capable engineering agents.