A team led by WINLAB Faculty Member Jorge Ortiz wins Best Paper Award at ACM Conference on AI and Agentic Systems (CAIS 2026). This work is part of project TraceFix which develops a verification and repair system for safer multi-agent AI workflows
Professor Jorge Ortiz and his team have developed TraceFix, a system for finding and repairing coordination errors in agentic AI systems. The work addresses a growing problem in AI deployments: as language-model agents are increasingly used to coordinate tools, data sources, and other agents, their behavior is often governed by informal protocols that are difficult to inspect, test, or enforce. The paper titled “TraceFix: Repairing Agent Coordination Protocols with TLA+ Counterexamples” co-authored by Shuren Xia, Qiwei Li, Taqiya Ehsan, and Jorge Ortiz received a Best Paper Award (Outstanding Solutions Paper) at the ACM Conference on AI and Agentic Systems (CAIS 2026).
TraceFix is part of a broader research effort in the Sensing and Reasoning Lab and WINLAB to develop auditable, verifiable, and secure AI systems for real-world settings. TraceFix treats agent coordination as a protocol-level problem. Rather than relying only on prompt engineering or post-hoc evaluation, the system analyzes the structure of an agent workflow, checks whether the workflow can violate expected coordination rules, and produces repairs when those violations are found. The approach brings ideas from formal methods and runtime verification into the design of AI agent systems, where failures may arise not from a single bad model output, but from the interaction among agents, tools, messages, and state transitions.
As agentic AI moves from demonstrations into operational environments, coordination bugs can have serious consequences. Agents may call tools in the wrong order, skip required approvals, expose information to the wrong component, or continue execution after a condition should have halted the workflow. These failures are difficult to catch with standard testing because they often depend on long interaction traces and rare combinations of events.
TraceFix provides a principled way to model these behaviors, search for counterexamples, and guide repairs. The system can identify protocol violations, generate diagnostic traces that explain how a failure occurs, and support changes that make the workflow satisfy the intended coordination policy. This gives developers a more reliable path from informal agent designs to systems whose behavior can be checked before deployment.
The work has applications in domains where AI agents must operate under explicit policies, including cybersecurity, data access control, scientific workflows, enterprise automation, and critical infrastructure. In these settings, organizations need more than high average performance; they need evidence that an AI system follows the rules it was designed to follow.
Further discussion of this work can be found on the project page: https://ortiz.rutgers.edu/blog/tracefix/