Agentic AI
Study exposes gap between AI coding agent promises and actual accountability records
Research reveals that responsibility frameworks for autonomous coding agents operate on two disconnected layers—platform permissions and vendor liability terms—leaving no clear audit trail when they conflict.
As AI coding agents grow more autonomous—authoring commits, opening pull requests, and deploying to production—a fundamental accountability blind spot has emerged. A new study examining how responsibility flows through agentic software development workflows reveals that the rules governing what these tools can do and who answers for their output operate in separate silos, sometimes contradicting each other with little to show for it.
How the system is supposed to work#
Responsibility for autonomous coding agents currently lives in two layers. First, platform controls determine which actions an agent is technically permitted to perform—what it can merge, deploy, or publish. Second, vendor liability terms spell out who bears legal and operational responsibility for the output. In theory, these work together. In practice, they often don't align.
The research approach#
Researchers conducted a detailed document study of 121 archived items, including documentation for four agentic coding tools, the platform controls restricting their operations, and eighteen policy documents from seven providers. They analyzed the material against five predefined dimensions—authority, execution, verification, consequence, and record—across nine key workflow events where conflicts might occur.
Key findings: where the mismatch appears#
The analysis surfaced several critical gaps:
-
Cursor's terms require users to verify suggestions, but its documentation describes a product that performs that verification automatically and records it as approval. No evidence trail exists showing the user actually discharged this duty.
-
At merge points, tools diverge wildly: two require a human to sign off, one documents an agent approving below a risk threshold, and one merely advises. Of fifteen mechanisms that compel verification, six rely on vendor-specified defaults; nine use defaults the study had to infer from other sources.
-
Absence claims break down: when vendors stated that certain safeguards existed, readers who weren't involved in creating them disputed five out of eight claims. Three claims held up under scrutiny, though one required material qualification.
