Method · Cross-impact & morphological analysis
Combinatorial explorers
Cross-impact & morphological analysis, reimagined for the agent economy
Exploring how many drivers combine and interact — across a space too large for any human team to do more than sample its corners.
The classical method
Two classic techniques tackle complexity head-on. Morphological analysis — developed by Fritz Zwicky and formalised for futures work by Tom Ritchey's General Morphological Analysis — breaks a problem into its dimensions and systematically explores every combination of their values, discarding the ones that are internally inconsistent. Cross-impact analysis studies how events influence each other's likelihood — how one development makes another more or less probable. Both fight the same human tendency: to consider only a few familiar combinations and call it a survey.
The shared premise is that the future is not a list of independent variables but a space of configurations, where drivers reinforce, cancel and gate one another. The interesting futures are specific, coherent combinations — and the dangerous ones are the combinations nobody thought to assemble.
Why the human-run version hit a ceiling
The blocker has always been combinatorics. A handful of dimensions with a few values each explodes into thousands of combinations, far beyond what a workshop can examine — so in practice morphological analysis becomes a sampling exercise that inspects the dozen or so configurations the group happened to think of, and cross-impact matrices stay small enough to fill in by hand. The space gets sampled at its obvious corners, and the coherent-but-overlooked configurations in the middle are never visited.
How it works with agents
Combinatorial-explorer agents turn the sampling exercise into a genuine search. They traverse the whole configuration space, discard the internally inconsistent combinations, model how each driver shifts the probability of the others, and surface the configurations that are coherent, plausible and overlooked — the futures that sit in the part of the space no human happened to point at.
The cross-impact half is handled at the same scale: rather than a hand-filled matrix of a few events, the agents reason over how a large set of developments raise and lower one another's likelihood, finding the reinforcing loops and the tipping combinations where several drivers move together. What was a back-of-envelope influence matrix becomes a navigable model of how the drivers actually interact.
The shift is qualitative, not just faster: the agents can cover the space rather than its corners. The output is not 'here are the three scenarios we sketched' but 'here are the coherent configurations across the whole space, ranked, with the inconsistent ones already removed and the interactions modelled'.
What the evidence says
This is the most purely computational of the methods, and the case for automating it is almost arithmetic: the value comes from exhaustively traversing and consistency-filtering a combinatorial space, which is exactly what machines do and humans cannot. Ritchey's own General Morphological Analysis work frames the method around cross-consistency assessment precisely because the combination count defeats unaided analysis — the agent simply removes the ceiling that forced the sampling compromise.
The generative-novelty evidence (Wang et al.'s SciMON) matters here too, because the prize in morphological work is the coherent configuration nobody anticipated — and an engine that can be pushed past the familiar is what surfaces it. The honest limit is that 'coherent and plausible' is not 'likely'; the explorer maps which configurations hang together, and human judgement and the forecasting agents supply how probable each one is.
Applied to the agent economy
The agent economy has many interacting dimensions — capability, autonomy, regulation, adoption, ownership, trust, concentration — and they do not move independently: tight regulation changes adoption, concentration changes trust, a capability jump changes all of them. Combinatorial explorers map how those drivers combine and reinforce one another, finding the coherent futures and the tipping combinations that a 'pick three scenarios' approach would skip entirely.
It is the natural feeder for scenario planning: the explorer surveys the whole space and the coverage map shows what was found, then the scenario engine narrates the most consequential coherent configurations and the human workshop chooses which to take seriously. The combination that would have been missed is exactly the one this method exists to catch.
Where humans stay in command
The failure mode is false coherence at scale — an engine that can assemble thousands of configurations can also assert that an impossible or contradictory one hangs together, and do it thousands of times. So cross-consistency filtering is explicit and auditable, every surviving configuration carries the reasoning for why its drivers are compatible, and a red-team agent attacks the most consequential configurations for hidden inconsistency before they feed downstream.
And probability and selection stay human. The explorer's job is coverage — making sure no coherent configuration is missed; deciding which configurations are likely enough and important enough to build strategy around is a judgement that belongs to the forecasting agents under human control and, finally, to the analysts who own the call.
How we run it
- Decompose — break the question into its key dimensions and their possible values.
- Enumerate — agents traverse the full combination space rather than sampling its corners.
- Filter — internally inconsistent or impossible combinations are removed by explicit, auditable cross-consistency assessment.
- Model — cross-impacts are computed, surfacing reinforcing loops and tipping combinations, with the most consequential configurations attacked by a red-team agent.
- Hand off — the coherent, high-impact configurations feed the scenario engine and the forecasting agents for narration and probability.