Every team running agents has this transcript somewhere. The agent "did it wrong." Someone pulls up the session to see where the reasoning broke — and finds that it didn't. The agent did exactly what it was handed. The failure happened before the run started, at the moment the work changed hands.
Here is the mechanism, plainly. Work moving through a human-plus-agents system crosses seams — human to agent, agent to agent, agent back to human. A seam transmits only what is made explicit. State does not travel by telepathy between sessions, tools, or teammates. Whatever was in your head, in a prior chat, or in a file nobody pointed at did not cross. The next worker starts with what arrived, not with what existed.
Which means every seam needs three things, written down.
The artifact. What exactly is being passed. A file, a brief, a draft with a path. Never "the context" — if you cannot name the thing that crossed, nothing crossed.
The owner. Who acts next. One name, human or agent, singular. "The team" is not an owner, and neither is "the agents." Work with two owners has zero.
The check. The definition of done the next worker can test without asking. A check that requires a follow-up question is not a check. It is a second handoff, unscheduled.
There was a clean public example of this working recently. One of the major browser teams reported that AI-assisted scanning helped find and fix a large batch of security bugs across recent releases. The interesting part is not the batch size. It is the shape of the loop: the agent did the exhaustive sweep no human team could afford, a human judged what the sweep surfaced, and every fix shipped through the normal release gate. Hundreds of findings crossed a human-agent seam without a pileup — because each one arrived as an artifact, with an owner, against a check. The scale came from the machine. The judgment never moved.
A while back I argued that nobody ships one agent — they ship a system. Later, that a longer leash matters less than the checkpoint. This is the third piece, and the one that holds the other two together: the checkpoint governs permission; the handoff governs continuity. A checkpoint asks whether this may proceed. A handoff asks whether the next worker has what it needs. Different mechanism, different failure — and the second one is the one hiding inside most "the AI messed up" stories.
So here is the Monday move. Pick one workflow you actually run. Mark every place the work changes hands — person to agent, agent to agent, agent back to person. For each seam, write one line: artifact, owner, check. Any seam where you cannot fill all three is not a mystery waiting to happen. It is your next AI failure, already scheduled.
Takeaway: A system is a promise that work survives changing hands. Spec the seam — artifact, owner, check — and most of your "agent errors" stop being errors at all. ✱
