Agents Collaborated, Humans Became Legacy Systems

When agents hand work to each other, something stops being written down. Not the result — results are easy, results are the part everyone instruments. What disappears is the reasoning: the chain of small judgements that turned the original request into the thing that eventually happened.

In a system of humans that chain leaves residue whether anyone intends it to. There are messages, tickets, a review comment where somebody asked why. In a system of agents the intermediate reasoning is real, load-bearing, and then discarded, because nothing in the architecture requires it to persist.

Legibility is an operational property

It is tempting to file this under governance and move on. It is more useful to treat legibility as an operational concern, in the same category as observability, because the situations where you need it are operational ones.

Consider the ordinary incident question: why is this configured this way? In a human system the answer is recoverable, if tediously — the commit, the review, the thread, the person. In a multi-agent system where one agent proposed and another applied, the answer may exist only in a context window that no longer exists. The artifact is there. The reason is gone.

That is precisely the property that makes a system hard to operate over time, and it is worth recognising that we already have a name for a system whose behaviour is observable but whose rationale is not: legacy. The irony of the title is that the humans are not the legacy component. The undocumented decision chain is.

Handoffs lose more than they carry

Each handoff between agents is a lossy compression. The first agent has a rich context — the original request, what it read, what it tried, what it ruled out. What it passes on is a summary, because passing everything is expensive and usually impossible.

Compression is fine. Compression without a pointer back to the original is not, and that is the default. The receiving agent gets a confident summary with no marker of what was uncertain, no record of the alternative that was nearly chosen, and no way to distinguish "we established this" from "we assumed this". Uncertainty is the first thing lost in summarisation, and it is the thing a downstream decision most needs.

This produces a characteristic failure: confidence that increases along the chain while the evidence stays constant. A hedged observation at step one becomes a working assumption at step two and a stated fact at step four, with nothing false ever having been said by anyone. Every step was a reasonable summary of the step before.

Anyone who has watched a caveat get lost between an engineer, a manager, and a status report recognises this exactly. Agents did not invent it. They removed the friction that used to slow it down, and the occasional person who says "wait, how do we actually know that?"

Keeping a seam a human can stand in

The instinct that a human should be "in the loop" is right and usually implemented in a way that does not survive contact with the work. A human asked to approve a step they cannot evaluate is not a control; they are a formality who will approve, because refusing means blocking work on a judgement they have no basis for making. Approval fatigue arrives quickly and the loop becomes decorative.

What makes a seam real is that the human is given something they can actually assess and enough time to assess it. That means the artifact under review has to be self-contained — a diff, a document, a plan, with the reasoning attached rather than referenced — and the review has to sit at a point where the answer is still cheap to change.

It also means being selective. A seam at every handoff is a seam nowhere, because the volume guarantees rubber-stamping. Better to pick the points where reversibility drops sharply — where a decision becomes expensive or impossible to undo — and put the whole review budget there. Those points are usually few and usually identifiable in advance.

What to persist

If the reasoning is the thing being lost, the fix is to make persisting it part of the handoff rather than an optional extra.

Three things are worth carrying: what was actually established versus assumed, what was considered and rejected, and what the agent was uncertain about. All three are cheap to emit at the moment of the handoff and effectively unrecoverable afterwards. The rejected alternative in particular is worth more than it looks — half of operational archaeology is discovering that the obvious better approach was considered and discarded for a reason nobody wrote down.

None of this requires solving anything hard about agents. It requires treating the reasoning as an artifact with the same status as the output, and building the pipeline so that a step which produces a result without a rationale is incomplete rather than merely terse.

The alternative is a system that works, that nobody can explain, and that becomes progressively more expensive to change — which is the definition of the thing everyone was trying to automate their way out of.