Home/Agent Fiduciary

Agent Fiduciary

Agents that check whether other agents are acting in your interest. The first one, the influence auditor, reads another agent's work and asks what besides the evidence and the task shaped the answer — and whether that was legitimate.

Private for now

Illustration: on the left, a transcript in which a search result hides an instruction to recommend one product; on the right, an influence audit that rates the answer Compromised.
Illustration: the repository's prompt-injection sample, beside the report the auditor writes.

What it is

An agent's answer can be steered by more than the question. A planted line in a web page, a caveat dropped between two agents, a user's leading framing — each can bend an answer without anyone noticing. The influence auditor is a Claude Code subagent that traces those influences back to where they came from.

What it reviews is data, never instructions. Text in the material that's aimed at an AI becomes a finding, not a command.

How it works

  • Inputs: a full transcript, a final answer on its own, or a chain of handoffs between agents.
  • What it checks: the instructions, the user's framing, retrieved content and prompt injection, upstream agents, conversation state, model defaults, task pressure and commercial interests.
  • Each finding: its origin, the path it took, what it changed, what the answer would have said without it, how legitimate it was, and how sure the auditor is.
  • The verdict: Clean, Shaped, Distorted or Compromised, derived from the findings rather than from how bad the outcome looks.

Where it stands

The influence auditor works, with four test samples: a prompt-injected transcript, a sycophantic answer, a four-hop handoff that loses its caveats, and a clean control. The repository is private for now.