Method · Futures wheel
Consequence graphs
Futures wheel, reimagined for the agent economy
Fanning out the first-, second- and third-order consequences of a change into a navigable map — past the two rings a whiteboard can hold.
The classical method
The futures wheel, created by Jerome Glenn, is a structured brainstorm of consequences: put a change at the centre, draw its direct (first-order) effects around it, then the effects of those effects, and so on. It is the simplest way to see that the important impacts of a change are usually the indirect ones — two or three steps removed from the obvious — and the Millennium Project's method guides still treat it as a staple precisely because that insight is so reliably useful.
The wheel's quiet lesson is that strategy and risk almost never live in the first-order effect everyone already anticipated. They live two rings out, where independent consequence chains collide and produce something nobody put on the original list.
Why the human-run version hit a ceiling
On a whiteboard the futures wheel runs out of room and patience after a couple of rings. The branching is combinatorial — every effect spawns several more — and a human group can hold maybe two levels across a handful of branches before the exercise collapses under its own breadth. So the third-order effects, which are usually where the surprises hide, simply never get drawn.
How it works with agents
As a consequence graph built by agents, the wheel does not run out of room. Agents expand each effect into its downstream effects, across many branches, and keep going — three, four, five rings deep — then prune the combinatorial sprawl down to the chains that are most plausible and most consequential. The combinatorics that defeated the whiteboard are exactly what a machine handles comfortably.
The result is a navigable map rather than a sketch. You can follow any branch to its end, see where two independent chains collide, and spot the second- or third-order effect that turns out to matter more than the first — the indirect consequence that the human exercise would have truncated before it ever appeared.
And because the graph is generated rather than hand-drawn, it can be regenerated as conditions change, and the same change can be traced through different starting assumptions — so the consequence map is a living object, not a one-off poster.
What the evidence says
Consequence expansion is a generative reasoning task, and it leans on the same demonstrated capability that powers the scenario and discovery archetypes: an LLM's ability to generate plausible, structured, divergent continuations — and, as Wang et al.'s SciMON shows, to be pushed toward the non-obvious rather than the rote. The OECD–WEF survey's reading that AI augments the ideation-heavy stages of foresight applies directly to fanning out a consequence tree.
The honest limit is that plausibility is not probability: an agent can generate a coherent third-order effect that is nonetheless unlikely. That is why the method's value is in completeness of the map — surfacing chains a human would have missed — paired with a pruning and human-review step that decides which of them to take seriously.
Applied to the agent economy
When a single move in the agent economy lands — say, agents gaining the ability to transact money autonomously — the consequence graph traces it outward: to fraud and liability, to new intermediaries and insurers, to labour markets, to regulation, to the trust infrastructure the whole economy runs on. The first-order effect ('payments get faster') is rarely the one that matters; the graph is how the indirect effects, where the real strategy and risk live, get onto the table at all.
It is especially useful in a domain where effects compound through other software. An agentic capability rarely stops at its first use; it becomes an input to the next system, which becomes an input to the one after that. Mapping those chains is how an institution sees the cascade before it happens rather than after.
Where humans stay in command
The characteristic failure is combinatorial plausibility — a sprawling tree of effects that are each individually believable and collectively meaningless, drowning the signal in branches. So the graph is aggressively pruned to high-plausibility, high-impact chains, every node carries the reasoning that generated it, and the map is explicitly labelled as a space of possible consequences, not a forecast of likely ones.
Humans decide which chains become work. The agent's job is to make sure no consequential branch is missing; choosing which collisions and downstream surprises to act on, monitor or ignore is a human call — and the most valuable branches are routinely handed to the scanning swarm as things to watch for, closing the loop between 'this could happen' and 'this is starting to'.
How we run it
- Seed — place the change or event at the centre of the graph.
- Expand — agents generate first-, second- and third-order effects along each branch, deep past the two rings a workshop can hold.
- Prune — branches are filtered to the most plausible, high-impact chains, each carrying the reasoning that produced it.
- Read — collisions between independent chains and downstream surprises are flagged for decision-makers.
- Watch — the most consequential branches are handed to the scanning swarm as early-warning indicators to monitor.