Agent accountability is moving from 'later' to 'now': deployers, not vendors, will own the liability

No-Conflict4823 · reddit · 2026-10-09

A Reddit deep-dive argues accountability for AI agent actions is arriving fast: underwriters are pricing agent risk, auditors demand evidence, and customers ask who signed off. Coding agents pushing to prod, support agents issuing refunds, ops agents changing infra and browser agents acting in SaaS all hold delegated authority—"the model did it" won't hold.

Key positions:

Name the accountable principal before execution, not in the postmortem; builder, deployer and approver are distinct roles with distinct liability, and most orgs haven't separated them.
Approvals must bind to the exact action and parameters (contract or hash), not to a drifting plan or session.
Evidence must survive disputes: mutable logs are reconstruction, not proof—tamper-evident, externally anchored records attributing agent identity, input, policy version and approver are the bar auditors and insurers will demand.
Vendor contracts mostly don't cover agent-initiated actions, so liability and controls land on whoever deployed the agent.

The author closes with open questions for practitioners: how to split ownership across the three roles, what evidence auditors accepted or rejected, agent-specific insurance sublimits, real time-to-answer on incidents, and whether standards land on per-action attestation or per-scope grants.

Original post →

Agent accountability is moving from 'later' to 'now': deployers, not vendors, will own the liability

More from coding & agent