Nothing Is Ever Done

In the U.S. Treasury market, a trade isn't finished when both sides agree on a price. A clearing process sits between execution and finality: matching what each party submitted, netting obligations, coordinating the actual movement of securities and cash. Under the Fed's operating rules, a posted transfer is final. A later reversal doesn't erase the original. The reversal is a new transfer, separately recorded. Participants who don't report an error within ten calendar days lose the right to contest it.

So "final" there doesn't mean nothing can ever change. It means change has rules: named actors, deadlines, a record of who did what and when. Finance calls the arrangement settlement, the point at which reopening a decision stops being arbitrary and becomes governed. The gap between "trade done" and "matter settled" is deliberate. Somebody priced out the cost of leaving it unmanaged and built machinery instead.

An agent finishes a task. It has sent a customer communication, updated a configuration, triggered a procurement workflow, changed an access policy. The system logs the task complete. The run has terminated. Is there a window in which that communication can be recalled, and does anyone own that decision? Is there a deadline past which unwinding the workflow becomes a different process with different signatures on it?

In most organizations deploying agents right now, no. Complete is the only status there is. Nothing sits between the agent stopping and the outcome becoming organizationally permanent. Permanent not because somebody decided it should be, but because nobody built anything else.

I've spent most of my career in a duller version of this. In infrastructure it goes: a deploy goes out, something downstream breaks, and at 3 a.m. you discover that rollback doesn't restore the state you assumed it did. Reversing local state doesn't retrieve the information, the expectations, or the choices the action already spent. Agent systems get the harder version, because the commitments have counterparties. A message lands in front of a person. A workflow starts somebody else's process, and a policy change gates somebody else's access. You can revert a row in a database and nobody is the wiser. You cannot revert the fact that a customer read something.

And agents produce these commitments quickly, faster than any approval chain or review cycle an organization currently has for deciding whether they should stand. Each task looks fine on its own. The dashboard is green. Underneath, commitments accumulate with no settlement structure, which means every completed task is either irrevocable by accident or reversible with no rules about who may do it.

A Sinch survey from earlier this year found that 74% of organizations running live customer-communications agents had rolled back or shut down an entire deployment over governance failures. That is the bluntest possible instrument: you kill the whole agent because you cannot govern a single outcome. It's what's left when there was never a clearing process and stopping trading is the only lever in reach.

The alternative is to build the thing finance built. Who is permitted to reopen an outcome an agent created? On what grounds, inside what window? And what is owed to everyone affected when somebody does?

My bet is that production readiness for agent systems turns out to have less to do with how often agents complete tasks correctly and more to do with whether an organization can answer those questions. A success rate is a fact about the agent. It tells you nothing about whether you can live with what the agent committed you to.