An accountable agent is not a more cautious model. It is an agent whose actions pass through a control path that produces a record. The pattern is the same whether the agent drafts a document or moves money, and it has four parts.
The control path
Propose, authorize, execute, record. The agent proposes an action. The action is checked against active policy and either permitted, denied with a cited reason, or escalated to a person. Only an authorized action executes. The full decision, including inputs, policy, and outcome, is appended to the record.
Least privilege, enforced at decision time
An agent's allowed actions are always a subset of the authority held by the person it acts for. The agent cannot widen that set, and any action outside it is refused and names the authority it would have required. This is least privilege checked at the moment of execution, not asserted in a document.
Delegation without surrendering control
To avoid approving every routine action, a person issues a standing policy: scoped to a class of actions, conditional, time-bounded, and revocable in one act. Revocation stops future automation immediately. The person who issued the policy remains the accountable party throughout.
A record that proves it
Each action is recorded with human contribution, AI contribution, the authorizing policy, the evidence, and the outcome, hash-chained so any later alteration is detected. The record is portable and verifies offline, so the agent's behavior can be proven to a third party years later without trusting the system that produced it.