What it looks at
AI agents are language models that act in steps. They plan, use tools, write code and hand work to each other. When several work together, they fail often, and the logs they leave behind are long and hard to read.
In 2025 Ezell, Roberts-Gaal and Chan proposed a framework for analysing incidents involving a single AI agent. This paper extends it to systems of several agents, using the layered structure of HFACS: what went wrong at the point of failure, and what sat behind it in supervision and coordination.
What it found
- The extended framework, AFACS, adds a supervision layer, a category for breakdowns in coordination between agents, and a way of coding violations from what an agent did, with the strength of the evidence stated each time.
- It was tested on MAST, a public dataset of multi-agent traces. An audit of that dataset found duplicate and corrupt rows, and logs that record what the agents did but none of the system documentation an investigator would need to reconstruct a failure.
- In a pilot, Claude coded 250 failed traces three times and applied the layers consistently. Supervision failures appeared in about two-thirds of failures, coordination breakdowns in about two-fifths and violations in 9–13%.
- Where the AI coder blamed the system’s design, it usually found a real weakness but overstated it. Design causes need design evidence.
Why it matters
Aviation learned that the cause of an accident is rarely found at the sharp end alone. The same holds for AI systems: a failure that looks like one agent’s mistake often rests on weak supervision or a poor handover between agents. The pilot shows the framework can be applied consistently. Whether the codes are right is the next study, with human coders.
About this paper
This is a working paper, version v1.0 of 4 November 2026. It has not been peer reviewed. Every reference was checked against its source, and the paper was audited by a reviewer that did not write it. As the paper says, the literature search, drafting and checking were done with AI assistance under the author’s direction, and the author is responsible for the text. Its sources, audit trail and, where there is one, its code and data are on GitHub.
