Skip to content

ADR-0011: Agent May Perform Explicitly-Instructed Status/Approval Edits

Status

approved — ratified by the Founder on 2026-07-11.

Date

2026-07-11

Context

The repository's hard prohibition against an AI agent marking a document status: approved (or an ADR accepted) was written broadly enough to read as barring the agent from ever making that file edit, even when a human with decision rights is present in an interactive session and explicitly directs the specific edit. While working through a backlog of items awaiting ratification, the Founder asked to have this clarified: the rule's actual intent is to stop the agent from deciding to approve something on its own, not to stop it from executing an edit the human has just explicitly authorized for that specific document.

Decision

The agent may perform the mechanical edit of a document's status field to approved (or an ADR's to accepted) only when a human holding the appropriate approval role (per decision-rights.md) explicitly and specifically instructs it to do so, for that named document, within that interactive session. The agent must still name the document and the status change being made so the human can confirm. The agent may never infer this authorization from a general instruction to "proceed" or "keep going," and a prior approval of one document does not carry over to another. The agent must never choose on its own that a document is ready to be approved.

Consequences

  • Removes friction for routine ratifications during interactive sessions with the Founder, where previously the agent could prepare everything but had to stop short of the final file edit.
  • Preserves the core safeguard: the agent still cannot self-certify its own proposals, and cannot treat a standing/blanket instruction as covering approval decisions.
  • Slightly increases reliance on the agent correctly recognizing when an instruction is specific and explicit enough to qualify, versus general. Ambiguous cases should still be escalated per escalation-and-ambiguity.md.

Alternatives Considered

  • Leave the rule as an absolute bar on any agent-performed status edit. Rejected because it forces the human to perform a purely mechanical file edit themselves even after they have clearly made the decision, adding friction without a corresponding safety benefit.
  • Let the agent infer approval from context or a general "proceed" instruction. Rejected — this reintroduces the exact risk the original rule guarded against (the agent self-certifying).

Follow-up Actions

Report an issue about this page