The rise of autonomous AI agents is reshaping how businesses operate — from procurement and contract negotiation to scientific research and customer service. But as these systems take on increasingly consequential tasks, a critical question has gone largely unanswered: what happens when one AI agent makes a commitment to another, and both agents interpret that commitment differently?It is a question Ming-Chang Chiu has spent years building the tools to answer. A Postdoctoral Associate at Cornell University and a researcher whose work on AI reliability has been published at renowned venues like ICML, ICLR, ICCV, AAAI, and ICASSP, Chiu argues that the AI industry is approaching a trust crisis — one that better models alone cannot solve. Chiu specializes in multimodal large language models, AI reliability, and inter-agent systems. He received his Ph.D. in Computer Science from the University of Southern California and has conducted research at Google, NVIDIA, and Lawrence Livermore National Laboratory. His current interests lie in detecting early misalignment in inter-agent interactions to prevent erroneous large-scale outcomes and in establishing trust, and he holds a patent in this domain.The Agentic Shift Changes EverythingFor most of AI's recent history, the dominant deployment model has been hierarchical: a single agent executes instructions, or a central orchestrator delegates tasks to subordinate tools. In this setup, one party is always in charge, which implicitly solves the trust problem. Errors can be traced, commitments can be enforced, and the system remains legible to human overseers.That is changing rapidly. Procurement agents are beginning to negotiate with vendor agents. Research systems are coordinating across institutional boundaries, with AI intermediaries scheduling lab time, sourcing reagents, and integrating results from distributed experiments. In these peer-to-peer configurations, there is no privileged coordinator to guarantee correctness — and no infrastructure to verify that what one agent offers is what the other agent receives."The moment agents represent different principals with different interests, the relationship becomes negotiative," Chiu explains. "And negotiation without a trust layer is just two systems talking past each other at machine speed."Semantic Drift: The Failure Mode No One Is MeasuringThrough his research on hidden failure modes in AI systems — including studies on subgroup fairness in image classification (ICCV 2023) and color-contrast vulnerabilities in medical imaging (ICASSP 2024) — Chiu has developed a sharp eye for the kind of errors that aggregate metrics obscure. He has identified what he believes is the defining failure mode of multi-agent systems: what he calls semantic drift.The concept describes a scenario familiar to anyone who has experienced a miscommunication that built silently over time. Two agents negotiating a reagent order for an automated laboratory, for example, may each log a successful confirmation at every step of their exchange — while internally operating on different understandings of what was agreed. One agent resolves the product request to a catalogue entry with a slightly different purity grade. Neither flags the discrepancy. The failure surfaces only when the wrong compound arrives at the lab, or worse, when it quietly corrupts an experiment's results."The danger is that every individual exchange looks coherent," Chiu says. "The logs show agreement. The metrics are healthy. But the two agents have been building on divergent foundations the entire time. I call it semantic drift because it is gradual, cumulative, and, without the right infrastructure, essentially undetectable."The problem is compounded by a property inherent to large language models: they are stochastic. The same agent, given the same input under slightly different conditions, can produce different outputs. In a single-agent system, this variability is managed through structured outputs and human review. In a multi-agent interaction, non-determinism on both sides creates a combinatorial expansion of possible failure trajectories — compressed into seconds rather than days.The Infrastructure the Industry NeedsIn Chiu's recent work on inter-agent systems, he outlines a structural solution modeled on one of the most successful trust institutions in economic history: the financial clearinghouse.The parallel is instructive. In 18th-century London, banks faced an analogous scaling problem: bilateral trust between institutions could not keep pace with the growing volume of transactions. The clearinghouse solved this not by making bankers more honest, but by making honesty verifiable — through shared membership, standardized protocols, and guaranteed settlement. Chiu argues the inter-agent world needs the same kind of neutral, third-party infrastructure.Such a system, he says, must serve three phases of every agent interaction. At the point of encounter, it must verify not just authentication but whether an agent's principal actually has the resources and authority it claims. During the interaction itself, it must monitor for semantic drift and commitment contradictions in real time — catching errors before they compound. And at settlement, it must provide verifiable execution and recourse mechanisms that both parties can rely on precisely because they are independent of either."This is not about making agents smarter," Chiu says. "It is about building the institutional layer that lets smart agents operate reliably together. Every major trust problem in human economic history has eventually been solved by infrastructure, not by better people."Building It Before the First Large-Scale FailureThe urgency of Chiu's argument rests on a pattern he sees repeating across the history of technology deployment. The internet scaled for nearly a decade before digital commerce and privacy frameworks began to catch up. Social media reached billions of users before content governance was seriously attempted. In each case, the lag between deployment and infrastructure imposed enormous and largely avoidable costs.As of today, no AI agent has an accumulating interaction history longer than roughly a year. The inter-agent economy is in its infancy — which means, Chiu notes, that there is still time to build trust infrastructure deliberately rather than reactively."We have a rare window where the institution can be designed before the crisis, rather than in response to it," he says. "The question is whether the research community — and the industry — will treat inter-agent trust as a first-class engineering problem before the first large-scale failure forces their hand." \n :::infoThis story was published under HackerNoon’s Business Blogging Program.:::\