The Agents You Forgot You Deployed

Governance cannot protect what it does not know exists

Ask most organizations how many automated agents are currently acting on their behalf, quietly pulling data, sending messages, approving requests, making small decisions at scale, and the honest answer is usually a shrug. Not because no one cares, but because the agents multiplied faster than anyone's inventory of them. A team spins one up to solve an afternoon's problem. It keeps running long after the problem is solved. Another gets copied from a template and modified just enough that nobody remembers it exists as its own thing. This is not a hypothetical failure mode. It is the ordinary condition of any organization that has adopted AI agents faster than it has built the discipline to track them.

Sprawl Is Not a Technical Failure. It Is a Stewardship Failure

The instinct is to treat this as an IT problem, something a better dashboard or a tighter deployment process will solve. That instinct is not wrong, but it stops short of the real issue. Every agent an organization deploys is a small delegation of judgment, a decision to let something other than a human being act, however narrowly, on behalf of the institution. Losing track of how many of those delegations exist is not merely inefficient. It is a failure of stewardship, the same kind of failure as losing track of who holds keys to a building or signing authority over an account.

Catholic Social Teaching has a name for the discipline this requires: subsidiarity paired with accountability. Authority delegated to a smaller or more local actor, in this case an automated one, does not dissolve the responsibility of the one who delegated it. If anything, delegation increases the obligation to know exactly what has been delegated, to whom or what, and with what limits. An agent nobody remembers deploying is authority exercised with no one answering for it, and that is precisely the condition every serious ethical framework for institutions warns against.

What Discovery Actually Requires

Recovering control does not begin with more sophisticated agents to police the less sophisticated ones. It begins with something far more basic: an honest, complete inventory, built and maintained as a standing discipline rather than a one-time cleanup project. Every agent needs an owner, a stated purpose, a defined scope of what it can touch, and an expiration point that someone actually revisits. None of this is glamorous. All of it is the digital equivalent of an institution knowing who has a key to which door, which is not a technical nicety but a basic condition of responsible governance.

The temptation, once an inventory exists, is to treat the discovery phase as the finish line. It is only the beginning. An agent's purpose can drift quietly from what it was authorized to do, especially when it continues to be modified by different people over time with no single owner tracking the cumulative change. Recovery, in the fullest sense, means periodically asking not just "does this agent still exist" but "does what it now does still match what it was trusted to do."

The Deeper Principle

Every technology that acts with any degree of autonomy on an institution's behalf inherits that institution's moral obligations along with its practical convenience. An organization that would never allow an untracked employee to make decisions on its behalf should not allow an untracked agent to do the same, however small each individual action appears. The scale of AI agents makes this discipline harder to maintain and, for exactly that reason, more urgent to insist on. Governance cannot protect what it does not know exists, and an institution's peace of mind about its AI agents is worth nothing if it rests on simply not having looked.