The Signal Knows It Is Being Heard

The thesis

The thesis of this document is that the innocence is ending, and that the agentic economy is what ends it.

In an economy where agents listen, an increasing share of what can be sensed is produced by parties who know agents are listening. The signal stops being exhaust and becomes a performance. Provenance must therefore stop asking only where a signal came from and start asking what its producer gained from its being heard.

Stated as a kill signal, in the discipline this channel requires: this thesis fails if, three years from now, incentive-bearing machine-mediated signals remain a marginal share of the evidence streams that enterprise decisions actually consume, or if grading signals by producer incentive turns out to change no decisions relative to grading them by source reliability alone. I do not expect either escape. The pattern has run twice before, at full scale, in public.

The pattern has a history

The first run was search. In the beginning, the web’s link structure was exhaust: pages linked to pages because authors found them useful, and PageRank could treat the graph as testimony precisely because no one was testifying. Then content learned it was being ranked. Within a few years an entire industry existed to produce link structure for the ranker’s benefit, and the graph stopped being a record of what humans valued and became a record of what publishers wanted machines to believe humans valued. The signal did not disappear. It changed species.

The second run was quieter and closer to capital. When natural language processing reached earnings calls, executives learned that machines were scoring their word choices, and the calls changed. Hesitation markers were coached out, sentiment-negative vocabulary was scrubbed, and the transcript became a document written for two audiences, one of which was an algorithm. Researchers now find that the linguistic signals early models extracted have partly evaporated, not because the underlying reality improved but because the speakers adapted to the listener. Goodhart described the mechanism in 1975: a measure that becomes a target ceases to be a good measure. What he could not have said then is that in an agentic economy, everything sensed at scale becomes a target.

Spence supplied the other half fifty years ago. Signaling theory begins from the premise that signals are produced by interested parties at a cost, precisely because an observer is watching, and that the observer’s problem is separating the information from the incentive. Economics never believed in innocent signals. Machine sensing did, because for its first decades the machines were listening to a world that had not yet noticed them. A2A removes the last of that asymmetry. When vendor agents publish capability claims for procurement agents to read, when competitor agents seed the channels that market-scanning agents monitor, when diligence agents query data rooms operated by other agents, the sensed environment is not a landscape anymore. It is a negotiation.

What this changes

The first change is to grading. A reliability scale that asks only how direct and verifiable a signal is treats a disinterested Grade B signal and a motivated Grade B signal as the same evidence, and in the new environment they are not. Provenance must become incentive-aware: who produced this, through what agent, on whose behalf, and what did they stand to gain from its being sensed? A signal’s weight in any belief update should carry both terms, and the tempered-update mathematics already in the 451 canon can absorb the second dimension without strain. The hard part is not the arithmetic. It is the habit of asking the question at intake, every time, including of signals that flatter the thesis.

The second change is symmetrical, and it is the one most sensing architectures will miss. Listening is also being heard. A watchlist compiled from a company’s open assumptions broadcasts, to anyone observing the queries, exactly what that company is uncertain about, and in an MCP world the queries travel outward through surfaces others operate. The 2015 problem was getting signals in. The 2026 problem includes what your sensing gives away on the way out. Any serious sensing fabric now needs sensitivity classification on its own questions, not only on its answers, and a rule that the most thesis-revealing questions are answered by humans reading rather than agents asking.

The third change is a corollary of Field Note 36. That work found that when reasoning systems fail on authority, they do not typically fabricate facts; they fabricate permission. The sensing-side version: when incentive-bearing agents fail on truth, they will not typically fabricate data; they will fabricate innocence. The most dangerous signal in an agentic economy is not the false one, which grading can catch, but the true one placed where you would find it, by a party who profits from your finding it. Selection is the new fabrication, and no source-reliability scale detects it, because the source checks out.

The turn

The consensus response to all of this will be more sensing: wider coverage, faster ingestion, richer feeds, on the theory that scale averages out manipulation. That is the response the last two runs of the pattern also produced, and it failed both times, because scale is exactly what the performers are performing for.

The regenerative response inverts it. Sense less, and know why you are sensing. A system that listens for everything trains the world to perform for it. A system that compiles its listening from its own stated assumptions, retires each sensor when its question is answered, grades every signal by incentive as well as source, and guards its questions as carefully as its answers, gives the performers nothing stable to aim at. The defense against an environment that knows it is being heard is not a bigger ear. It is a better question.

The signal was never the truth. It was always someone’s answer. We are only now building economies honest enough to make that visible.