Method · Scenario planning
Generative scenario engines
Scenario planning, reimagined for the agent economy
Building several distinct, internally-consistent futures for the agent economy — so no single bet goes unexamined, and the uncomfortable branch never gets quietly dropped.
The classical method
Scenario planning is the craft Shell made famous: instead of a single forecast, you construct a handful of plausible, internally-consistent stories about how the future could unfold, built around the critical uncertainties that matter most. It trains an organisation to be robust across several futures rather than right about one — which is why Pierre Wack's scenario teams left Shell better prepared for the 1973 oil shock than its competitors, and why the method outlived every point-forecast that was made alongside it.
The discipline is well codified — the 2×2 of critical uncertainties, the four archetypes (collapse, discipline, transformation, continued growth), the cone of plausibility — in toolkits from the UK Government Office for Science, the UNDP Foresight Manual and the EU JRC's Scenario Exploration System. What all of them protect against is the same human failing: the tendency to plan for one future, usually the comfortable extrapolation of the present, and to be ambushed by any of the others.
Why the human-run version hit a ceiling
The thing that has always rationed scenario work in government is not the writing — it is the throughput of focal questions. A proper scenario set means convening the right dozen experts for a two-day workshop, which a unit can afford perhaps three or four times a year. So scenario thinking is reserved for a tiny handful of the largest strategic questions and is structurally unavailable to the hundred smaller-but-still-consequential decisions a government makes. Worse, human panels under deadline and political pressure systematically under-produce the uncomfortable, low-probability-high-impact branch — the very scenario the exercise exists to surface.
How it works with agents
The hard, slow parts of scenario work — identifying driving forces, combining them into coherent worlds, writing each one out in vivid, internally-consistent detail — are exactly what agents do well. A generative scenario engine maps the critical uncertainties (often themselves surfaced by the scanning swarm), combines them morphologically into a span of distinct futures, and drafts each as a rich narrative with its own driving logic, winners, losers and early-warning indicators.
Because it is cheap to generate, the engine produces a genuine plurality of futures rather than one house view — and it can be instructed and evaluated for explicit divergence, deliberately generating the tail-risk branches that human workshops shy away from. It also produces a coverage map showing which combinations of uncertainties were and were not explored, so blind spots become visible rather than hidden.
This inverts the economics that rationed the method. Because the engine can produce a credible, divergent, well-sourced first draft overnight for the cost of compute, the human workshop stops being the bottleneck and becomes the refinement step applied to a draft that already exists. Scenario thinking can then be applied an order of magnitude more widely — a procurement officer weighing a fifteen-year commitment can have a scenario set this week, not next quarter.
What the evidence says
The case rests on two things. First, the practitioner signal: the OECD–WEF AI in Strategic Foresight survey reports that scenario-building and narrative generation are among the foresight tasks practitioners most expect AI to augment — precisely because divergent ideation and consistent narrative drafting are language tasks, and language is what these systems do. Second, the capability signal: the same novelty-optimisation demonstrated in Wang et al.'s SciMON — pushing a model toward the genuinely new rather than the familiar — is what makes an LLM a credible engine for the divergence half of scenario work, where the characteristic failure of human panels is groupthink and premature convergence.
The honest boundary matters as much as the capability. The engine is good at breadth, speed and consistency; it is not a substitute for the human judgement about which futures to take seriously. The evidence supports automating the draft, not the decision — and the operating model is built on exactly that split.
Applied to the agent economy
We use it to ask the question leaders most need and least have time for: what are the genuinely different ways the agent economy could go? A world of a few dominant agent platforms versus open agent markets; tight regulation versus laissez-faire; fast labour displacement versus slow augmentation; high trust versus a cascade of agent-driven failures that poisons it. Each scenario becomes a lens a strategy can be tested against.
And because the engine refreshes scenarios as new signals arrive, the set never goes stale between flagship exercises. In a domain moving as fast as the agent economy, a scenario set that is eighteen months old is not a foresight asset — it is a museum piece. A living engine keeps the worlds current.
Where humans stay in command
Two characteristic failures, two guardrails. The first is bland convergence — the engine regresses to the most probable, most familiar future and quietly drops the tail risks, defeating the whole purpose; the answer is to instruct and evaluate explicitly for divergence and to require the coverage map, so a human can see which uncertainty-combinations were skipped. The second is false coherence — a beautifully written scenario that is internally contradictory or rests on an impossibility; the answer is that a red-team agent attacks every scenario set for internal consistency and hidden assumptions before any human sees it.
Above all, scenario sets are inputs to human workshops, never replacements for them. The standing rule is that an agent-drafted set is never circulated as a finished product — it is either taken into a human workshop or explicitly stamped 'draft, un-reviewed'. The agent widens the option space; humans prune it, politically contextualise it, and decide which futures to inhabit. That choice is the one thing the engine is forbidden to make.
How we run it
- Frame — agents identify the critical uncertainties shaping the agent economy, often drawn straight from the scanning swarm's registry.
- Combine — uncertainties are crossed morphologically into a set of distinct, internally-consistent worlds, with a coverage map of what was and was not explored.
- Narrate — each world is drafted with its drivers, dynamics, winners, losers and early-warning indicators (handed back to the swarm to monitor).
- Attack — a red-team agent breaks each scenario for internal contradiction and hidden assumptions before any human reads it.
- Pressure — strategies are wind-tunnelled against every scenario in a human workshop to find what holds across all of them.