Same money as today: a central bank, dollars, taxes. We pulled every AI lever in the simulator, 2,100 runs in all, to see which ones move jobs, prices and who ends up with the gains.
Based on Barry Becker’s agent-based economy simulator · experiments and write-up by an AI analyst · October 2026
Six questions, one monetary regime
The simulator runs thousands of simulated households, hundreds of firms, a few banks, a government and a central bank, month by month. AI reaches the economy through two doors. Task automation lets firms do more of their work with software and, later, robots. AI agents, owned by households, sell compute to firms and raise their capacity. Their real cost keeps falling, with an optional energy step such as fusion power. Everything here keeps the app’s fiat money and central bank as they are.
Where the numbers come from
Runs 20 seeds per setting, 500 households, 50 firms, 3 banks, 240 months. Figures are medians; brackets give the 10th–90th percentile across seeds. “Year 20” is the average of the last 12 months.
Live app buttons open the deployed simulator with the same settings and 10 seeds. Press Run.
The baseline
Substantial AI preset, unless stated: it multiplies output by 6.2 in 20 years, against 1.4 with no AI.
App defaults for everything else: 20% income tax, government buys 20% of income, UBI paid from 25% of the AI slice of GDP, a balanced central bank.
01
Question one · Two doors
Replace or augment?
Setting compute cost decline to 0 shuts the task-automation door: software never gets cheaper than a worker, so firms don’t swap people for it. Household agents still sell compute to firms. That “agents only” economy grows 2.9× in 20 years with unemployment at 5.0%, close to the 4.6% with no AI at all. Open both doors (the Substantial preset) and output reaches 6.2×, but 47% of households are out of work.
A machinist in each world. In the agents-only economy she keeps her job; wages are still 40% of household income and the typical household consumes 2.4× what it did at the start. Under full automation she has nearly an even chance of being out of work, wages fall to 9% of household income, and the typical household still consumes more, 3.5×, paid for mostly by profits and the AI-funded UBI.
Agents that augment firms add growth without layoffs. Replacing tasks adds about twice as much again, and that is the part that costs jobs.
02
Question two · Policy
Policy can’t buy the jobs back
We tried a hawkish and a jobs-first central bank, fully flexible wages, hiring that responds strongly to wages, eight job applications a month instead of three, much cheaper agents, a higher capability ceiling and deficit spending. Year-20 unemployment stayed within 45.9–47.3%. In this model, firms hire for the human share of their output, so the jobs number follows how much work AI does, not the policy mix.
Three things do move it. Robots that arrive in year 20 instead of year 8 keep unemployment at 39%, and a later adoption midpoint (year 15) gives 44%. A UBI paying out the entire AI slice of GDP brings it to 34%, but at a price: inflation averages 15% a year and the real wage ends at 0.53× its starting level, against 1.16× in the baseline.
Here the unemployment rate is a measure of how far automation has gone. Policy decides how jobless households live, not how many there are.
03
Question three · Cheaper, better agents
Falling agent prices: a threshold, then a windfall
At the app’s default start price (4% of a monthly wage), agents are already cheap enough for firms to buy nearly all they can use. Halving the price every 6 months or every 20 years changes year-20 output by less than 1% (6.23× vs 6.27×). So we also started agents at half a wage. Now the price path decides everything. With a 20-year half-life, agents barely sell and output reaches 3.1×, the same as with no agents (3.1×). With a 2-year half-life it reaches 6.3×. A fusion-style energy step that removes 90% of the cost makes up for a slow decline: 3.4× without it, 6.2× with the step in year 3 and 5.5× when a 20-year half-life gets the step in year 5.
The worker who isn’t replaced by cheaper agents. Cheaper agents don’t take more jobs: unemployment is 47% with the slow price path and 47% with the fast one. The real wage barely moves either (1.16× vs 1.17×). What changes is the household around her: median consumption climbs from 1.84× to 3.42× and median income from 1.88× to 3.50×, because agent income, profits and the AI dividend all grow. Her paycheck is a smaller share of a bigger pie: 9% of household income instead of 19%.
A steep fall in agent prices, whether from steady progress or a fusion-style energy step, is what unlocks the agent economy. Past that point, extra cheapness adds little; what agents can do matters more.
04
Question four · Capability
The singularity case: raise the ceiling
The capability ceiling caps how much bought compute can multiply a firm’s capacity. At the default of 1, it can at most double it. Output scales with the ceiling: 4.7× at 0.5, 6.2× at 1, 9.4× at 2 and 15.6× at 4. Median consumption follows (2.7× to 8.4×), and so does the median household’s real wealth. Faster capability growth adds nothing once the ceiling binds: ceiling 4 with growth 1 instead of 0.25 gives 15.6×.
Our singularity mix combines ceiling 4, faster capability growth, a 1-year price half-life and a 90% energy step in year 3. It reaches 15.7× output. Prices drift down 0.9% a year even with the central bank aiming for 2% inflation, with no bank failures (0 in the median run). Unemployment stays at 47%, and wages fall to 4% of household income.
A more capable AI creates a far bigger economy without costing extra jobs. The question becomes who owns it.
05
Question five · Distribution
Who captures the gains?
Two levers decide where AI income lands. Ownership concentration sets how strongly AI-era profits skew toward high-skill households. UBI share pays a slice of AI output equally to every household. We crossed four ownership settings with five UBI shares under Substantial AI. Output hardly changes across the grid (6.2–6.8×). The median household changes a lot: from 2.6× its starting consumption with narrow ownership and no UBI to 6.1× with broad ownership and a full payout.
Broad ownership costs nothing in inflation: 4.4× at 0.7% a year, against 3.0× under the default skew with no UBI. UBI is cheap up to about half the AI slice (1.6% inflation at 50%), costs more at 75% (3.2%) and much more at 100% (15%). The best setting that stays near the 2% target is broad ownership with a 75% UBI: median consumption 5.7× [5.6–5.8], inflation 3.1%, and the poorest fifth holds 5.5% of wealth, against 2.2% at the defaults.
The singularity, shared or not. With the singularity mix and the default skew with no UBI, the median household consumes 7.1× its starting level while the richest tenth holds 57% of wealth. With broad ownership and a 75% UBI, the same economy gives the median household 13.8×; the richest tenth’s share drops to 37% and inflation is 2.2%.
How the government is run also matters. The app’s budget presets barely change output, but median consumption ranges from 2.6× under a small state to 4.6× under the AI-dividend preset (35% tax, 75% UBI, 4.3% inflation). A large state reaches 4.2× at 0.5% inflation, and deficit spending 4.4× at 2.8%. The central-bank presets change none of these outcomes much: median consumption is 3.50× hawkish, 3.52× jobs-first and 3.52× monetizing.
Spreading ownership is the inflation-free lever. UBI adds a lot until it approaches the whole AI slice; past that, inflation eats the real wage.
06
Question six · Business cycles
Recessions move from the jobs report to the dividend check
The simulator’s shocks hit demand, credit or productivity for two years: one of boom, one of bust. We compared each seed with and without shocks. With no AI, frequent large shocks add 45 months of shrinking output, 4.6 points to peak unemployment and 9.2 points to the worst fall in median consumption. Under Substantial AI, they add 24 months of shrinking output and just 0.2 points to peak unemployment, because the jobs number is already set by automation. The consumption hit doesn’t shrink: 10.7 points.
A household that lives on dividends. When profits make up most household income, a credit bust or demand slump shows up first in the profit check and the AI dividend, not in a layoff notice. With frequent large shocks, profits are 81% of household income under Substantial AI, so every downturn reaches the household budget even when the household had no job to lose.
Automation steadies output and employment through a downturn, but not household spending. The hit lands on incomes, not jobs.
What to keep in mind
Jobs follow a hiring rule. Firms hire for the human share of output, so unemployment tracks AI’s share. That is a modeling choice, and it is why the labor-market sliders barely move it.
Small economies. These runs use 500 households and 50 firms. Spot checks at the app’s default size (4,000 households, 200 firms, 4 banks, 4 seeds each) give 1.42× / 6.18× / 15.4× output for no AI / Substantial / singularity mix, against 1.43× / 6.24× / 15.7× here, and 46.9% vs 47.0% unemployment under Substantial. Broad ownership with a 75% UBI gives 5.70× median consumption at full size vs 5.70× here.
Presets and dials. AI levels and budget presets copy the app’s catalogue. Ownership concentration skews profits toward high-skill households; it does not decide who owns agents.
A teaching tool, not a forecast. Every number here describes the model’s assumptions, not the real world.
Try it yourself
Open a case, press Run, then move one AI lever (capability ceiling, agent cost half-life, ownership concentration or UBI share) and run again.