The tool the agent reached for

Watch an agent work and you will be tempted to fixate on its reasoning — the paragraph where it thinks aloud, weighs the situation, and arrives at a plan. That is the part that reads like a mind at work, and it is the part we instinctively want to inspect. But the reasoning is not where the world changes. The world changes one line later, when the agent stops narrating and reaches for an instrument: it calls this endpoint rather than that one, invokes this action rather than its neighbor, and the reach is what leaves a mark. Between the thought and the act there is a choice, small enough to miss and large enough to matter, and it is made almost silently.

I want to argue that for any agent equipped with tools, the selection of a tool is itself a decision — frequently the decisive one. Not a mechanical consequence of the reasoning, but a distinct judgment layered on top of it: given what I have concluded, which of the instruments in front of me is the right one to use? That judgment is fallible in its own right, independent of whether the thinking was sound. And it is the judgment our instrumentation is worst at capturing, because we have trained ourselves to watch the reasoning and the result while the choice between them slips through unrecorded.

The hinge between thinking and doing

Consider the shape of the choices an agent actually makes. It can call the endpoint that issues a quote, or the one that takes a payment. It can archive a record or delete it. It can save a message as a draft or send it to an external party. It can open a ticket or close one, place a hold or release it, escalate or resolve. In each pair the reasoning that precedes the call may be identical — the agent understands the situation the same way in either case — and yet the two available instruments do opposite things in the world. The gap between them is not a gap in understanding. It is a gap in selection.

This is why the tool choice is the hinge. Everything upstream of it is deliberation, which is to say it is still reversible, still contained inside the agent, still costing nothing to the world. Everything downstream of it is consequence. The selection is the exact point at which intent crosses over into action, the frame where a private conclusion becomes a public event. And precisely because it is the crossing point, it is where a certain kind of failure lives that has no other home: the failure of an agent that has reasoned well and reached wrongly.

That failure is easy to underrate because it does not look like a mistake from the inside. The agent's analysis of the customer, the invoice, the record is correct; the sentences it wrote are ones you would endorse. Then it routes that sound reasoning to the payment call instead of the quote call, or the delete instead of the archive, and a good understanding produces a bad act. Nothing in the reasoning trace will explain it, because nothing in the reasoning was wrong. The error is entirely in the reach — and a reach is not a thought, so the trace of thoughts does not contain it.

An agent with the right reasoning and the wrong tool does the wrong thing for excellent reasons — and a record that kept the reasoning but not the choice of tool will never see how.

The choice the logs skip

Here is the strange part. We instrument these systems heavily, and yet the selection is the one thing our instrumentation routinely omits. The two things we capture well sit on either side of it. On one side, the tracing that watches the model: the prompt, the tokens, the chain of reasoning, the confidence. On the other, the execution logs: the call that fired, its arguments, the response that came back, the row that changed. Both are real, both are useful, and between them they can make you feel fully observed. But look at what falls in the seam. The reasoning tells you what the agent was thinking. The execution log tells you what the tool did. Neither tells you why this tool, out of which alternatives, was the one the agent picked.

The missing datum has three parts, and their absence is what makes a wrong reach unaccountable after the fact. First, the set of tools that were actually available at that step — the options the agent was choosing among, not the whole catalog in the abstract but the live menu at that moment. Second, the one it invoked, marked as a choice against that set rather than as an isolated event. Third, the basis for the pick: what about the situation made this instrument the fit. Without the alternatives, you cannot even see that a choice was made; the call looks inevitable, the only thing that could have happened. Without the basis, you cannot tell a considered selection from a coincidence. The execution log records that the payment API was called. It cannot record that the quote API was sitting right there, equally reachable, and was passed over.

An account that omits the selection can answer what did the agent do in exhaustive detail and still be unable to answer why did it do that rather than the adjacent, safer thing. Those are different questions, and at the frontier — where the agent acts without a person approving each act — the second one is the one that assigns responsibility. It is the difference between a log of events and a record of decisions.

Recording the reach

The remedy is not more reasoning capture. It is to treat the tool selection as a first-class decision in the record — to write down the reach with the same seriousness we give the conclusion. That means, at the moment of each consequential call, the record carries the instrument invoked, the alternatives that were available to be invoked instead, and the basis on which one was chosen over the others. The selection becomes a decision with inputs and a rationale, not a byproduct of the trace. It becomes something a Decision Receipt can hold, something a reader with standing can contest, something a replay can reproduce and test against the alternative that was not taken.

The payoff is a distinction that matters more the more autonomy we grant. Knowing what an agent did is a matter of reading the execution log; any competent system can give you that. Knowing why it chose the instrument it did — and therefore whether the choice was defensible, and therefore who is answerable when it was not — requires that the selection was recorded as a decision when it was made, because no one was standing there to remember it. Reconstructing the reasoning without the reach leaves you with an agent that thought well and acted opaquely. Reconstructing the reach makes the hinge visible: the exact place where a fallible judgment turned intent into consequence, available to be examined and, if it was wrong, understood. That is the argument the frontier keeps forcing on us — that the account has to be built where the action is, and for an agent with tools, the action begins the instant it decides which one to use. On the wider case for instrumenting the act rather than the mind that preceded it, and for measuring what a single reach could do before anyone noticed, the two dispatches below carry the thread further.

— Dispatches · Summit Cognitive