Somewhere in your stack, one agent is about to accept another agent's work as fact.
It will not check it. It has no way to check it. What arrives is a finished thing — a summary, a shortlist, a spec — and finished things look identical whether they were verified or invented. So the second agent proceeds. The third accepts the second. Every step reports success, and the only red light in the whole chain is at the end, in a person's hands, with four green logs behind it.
Here is the mechanism, and it is arithmetic before it is anything else. Reliability multiplies along a chain. Four agents that each get it right nine times out of ten produce a chain that is right about two times in three. Nobody sets out to build a sixty-six percent system. They build four ninety-percent ones and quietly assume the nineties will hold hands.
Confidence, meanwhile, does not multiply. That asymmetry is what makes this expensive. Each hop reports a binary — done, complete, success — so the uncertainty is discarded at every boundary while the certainty travels at full strength. By the fourth hop the work is less reliable and more confident than it was at the first. The chain is not only failing. It is laundering.
This is no longer a thought experiment. Delegation between agents built by different teams, on different stacks, is becoming ordinary plumbing. The moment your agent can hand a task to one you did not build, the question stops being how good your agent is and becomes what your agent is willing to believe.
What is missing at each boundary is the warrant: the basis on which the previous step was accepted. Not the output, and not a log written for auditors after the fact. A short header that travels attached to the work and answers three things. What was checked. Against what condition. What was not checked.
That third field is the one that does the work. An output with no stated gaps invites the next worker to assume there are none. An output that says sources verified, pricing not verified has told the receiver exactly where to spend its scepticism, and a receiving agent can act on that — route it to a human, run one more check, or refuse the task outright. Uncertainty that is written down is routable. Uncertainty that is dropped becomes somebody's Thursday.
Three weeks ago I argued that every seam needs an artifact, an owner, and a check. That holds when you own both sides of the seam. This is the harder case: when the next worker is not yours, the check cannot live in your process. It has to ride along with the artifact, or it does not exist.
So the move is a subtraction before it is an addition. Draw the chain you actually run and count the hops. Multiply your honest per-hop success rate across them — not the demo rate, the Tuesday rate. Compare that number to the way the chain currently reports itself. Then either shorten the chain, or make every hop state what it did not check.
Takeaway: Confidence passes cleanly between agents. Reliability does not. Make each hop declare what it left unverified, or the last human in the line inherits every gap at once. ✱
