Method · Wild cards & weak signals
Anomaly hunters
Wild cards & weak signals, reimagined for the agent economy
Hunting the low-probability, high-impact events that trend-followers systematically miss — and the faint early traces that would tell you one is becoming real.
The classical method
Most analysis extrapolates trends — and so it is blind to discontinuities. Wild-cards-and-weak-signals work deliberately looks the other way: for the low-probability, high-impact events (the wild cards) and the faint early traces that might precede them (the weak signals). It is the discipline of taking the improbable seriously before it stops being improbable, and it is the part of foresight that protects an institution against being told, after the fact, that nobody could have seen it coming.
The method exists because human cognition is built the wrong way for it. We round low probabilities down to zero, pattern-match the new to the familiar, and treat the absence of a precedent as evidence of impossibility. Wild-card work is a structured corrective to all three reflexes.
Why the human-run version hit a ceiling
Humans are bad at this in a way that does not improve with effort: we are biased toward consensus and toward the trend, so the outlier that contradicts the prevailing story is exactly the signal a human team is most likely to dismiss — and a faint signal, by definition, is easy to skim past when attention is scarce. The result is that wild-card work, done by hand, tends to track only the wild cards that are already fashionable to worry about, and misses the genuinely off-trend ones.
How it works with agents
Anomaly-hunting agents are built to do the opposite of the human reflex. They are tuned to prize the outlier, the contradiction, the thing that does not fit the trend — and to treat a faint, strange signal as worth tracking precisely because it is faint and strange, rather than discarding it for lacking corroboration. Scanning continuously and without the human pull toward consensus, they keep a live watchlist of wild cards and the weak signals that would tell you one is becoming real.
Each candidate wild card is paired with the weak signals that would foreshadow it, so the watchlist is not a list of fears but an instrument: it specifies, in advance, what early traces to watch for, and escalates when those traces start to strengthen. The improbable event stops being a surprise and becomes a hypothesis under active surveillance.
Because generation is cheap, the agents can hold a far longer tail of low-probability scenarios under watch than a human team would ever justify spending attention on — which is the whole point, since the wild card that gets you is almost always one that did not make the short, fashionable list.
What the evidence says
The weak-signal half of this is directly demonstrated: work on weak-signal detection over the research corpus turns the foresight-theoretic notion of a 'weak signal' into a computable pipeline that flags emerging fronts from sparse, early evidence — exactly the faint-trace detection a human skims past. Wang et al.'s SciMON reinforces the capability by showing that novelty can be optimised for explicitly, which is the machine analogue of prizing the outlier.
The wild-card half is grounded in the risk literature rather than the forecasting literature. The International AI Safety Report 2026 is unambiguous that frontier AI systems carry genuine tail risks — capability jumps, emergent failure modes, hard-to-anticipate behaviour — which is precisely the class of low-probability, high-impact event this method exists to keep on the table. The agent's role is not to predict which wild card fires, but to ensure none of the plausible ones is being ignored.
Applied to the agent economy
The agent economy is a wild-card factory: a sudden capability jump that resets the board overnight, a cascading failure of interacting agents, an agent-driven market event, a security breakthrough or breakdown that no trend line predicted. Anomaly hunters keep these on the table — and watch for the early traces — so that a 'nobody saw it coming' event is instead one we were already tracking, with a considered position ready.
This is the method that most directly buys down the risk of strategic surprise, which in a domain moving this fast is the dominant risk. The institutions that get hurt are not the ones that guessed the wild card wrong; they are the ones that never had it on the list.
Where humans stay in command
The failure mode runs opposite to most of the fleet: not over-confidence but over-flagging — an instrument tuned to prize anomalies will manufacture alarm from noise if left unchecked, and a watchlist that cries wolf gets ignored exactly when it matters. So every wild card carries an explicit, low and honest probability, the weak signals attached to it are concrete and falsifiable, and escalation is triggered by traces actually strengthening rather than by the agent's standing unease.
And a human decides what to do with a strengthening signal. The agents keep the watch and raise the flag; judging whether a wild card has crossed from 'tracked' to 'act on it now', and what that action should be, is a human call — informed by the watchlist, owned by the analyst, and routed through the red-team before it reaches a decision-maker.
How we run it
- Hunt — agents seek outliers, contradictions and off-trend signals continuously, tuned against the human pull toward consensus.
- Hypothesise — each is linked to the specific wild card it might foreshadow, with concrete, falsifiable early-warning traces.
- Watch — weak signals are tracked over time, with honest low probabilities attached, for escalation.
- Warn — when the traces actually strengthen, the wild card is escalated early to an analyst.
- Decide — a human judges whether to act, monitor or stand down, after an adversarial review.