Confidence-Calibrated Knowledge Graphs Fix Multi-Agent AI Errors

Multi-agent AI systems are increasingly used to automate software engineering tasks like requirements analysis, design, and test generation. However, when agents operate sequentially over shared artifacts, low-confidence decisions from upstream agents propagate downstream, creating orphaned requirements, contradictory links, and compliance gaps—especially dangerous in safety-critical domains like automotive or aerospace. A new paper on arXiv (2606.17203) by Mohamed Essam and six co-authors introduces a trust-aware coordination framework to address this problem head-on.

The framework centers on a shared knowledge graph that serves as both semantic memory and a coordination surface. Agents assess and build upon each other's contributions using calibrated confidence scores. Key innovations include a two-stage traceability link prediction pipeline combining embedding-based retrieval with LLM-based multi-criteria analysis, a traceability seeding mechanism that compares derivation-time and validation-time confidence, and a consistency protocol with confidence threshold gating, divergence detection, and conflict resolution. Evaluations on an automotive software engineering case study demonstrate that confidence calibration is essential for effective pipeline coordination, with ablation studies confirming the protocol's impact on reducing error propagation.