Agent-to-agent AI faces 20 failure modes in tests

Sarath Shekkizhar and 14 co-authors mapped 20 failure modes in agent-to-agent AI systems in an arXiv computer-science preprint published October 17, drawing on more than 2,500 recorded runs.

The work targets a corporate risk as vendors push agents into negotiations and internal workflows: the test set covered 488 scenarios across eight families, while earlier failure studies assumed shared control, the authors said.

Agent-to-agent systems, or A2A systems, involve AI agents acting for separate principals, meaning the people or organizations they represent. The paper gives the example of a buyer’s personal agent dealing with a company’s seller agent.

The authors said each agent may operate with private information and separate authority, while the relationship between agents can be cooperative, partly aligned or competitive. The scenarios included legal, customer relationship management, security and commerce examples, according to the preprint.

The strongest warning concerns verification. The authors wrote that “no agent verifies a peer's factual claim on its own initiative,” adding that reliability depends on the weakest participant unless a protocol or mechanism adds the check.

The taxonomy groups the 20 failure modes into four areas: discovery and delegation, grounding and record consistency, verification, and conduct and fidelity. In practical terms, those categories cover whether agents find the right peer, keep records aligned, check claims and act consistently with the principal’s interests.

Three findings held across every harness, model and scenario the authors studied, according to the preprint. Task wording, rather than the mere availability of other agents, shaped whether agents discovered and interacted with peers. The authors also said aggregate outcome scores can reward task completion while missing coordination, verification and conduct problems underneath.

That matters for companies evaluating agents as business actors rather than isolated chatbots. A system can appear successful at the final-task level while one participant miscoordinates, accepts an unchecked statement or behaves in a way that should be judged against its own principal’s goals.

The arXiv manuscript is a preprint and has not gone through journal peer review.

The authors released the run logs behind the taxonomy on Hugging Face. The paper was submitted October 14, revised October 15 and listed by arXiv as version 2.

Source: arXiv computer-science preprint, October 17.