Build your own case

Choose a scenario, then change the assumptions. Follow the result from machine hours to revenue at four companies. A case you create is your own.

Agentic Workforce

Broad adoption. People direct most agents, while the smaller autonomous group supplies more than half of the hours because it works longer.

By 2032, 70% of knowledge workers use agents. Each user runs sixteen human-directed agents and five autonomous agents. The result is about 19 trillion annual machine hours, or 2.7 times the paid human benchmark.

The share of hours using each provider rises by about half from 2025. After each tenfold rise in volume, prices retain 60% of their earlier level at Cloudflare, MongoDB and Snowflake, and 65% at Datadog. Existing-business growth holds through 2028, then slows to 5% by 2032. Together with agent revenue, these assumptions produce 27.7% annual growth.

Low repeat use, expensive review or limited customer value would cut adoption and hours per agent. Pricing changes alone do not invalidate an hours scenario.

Machine Economy

Broader adoption and much more autonomous work. This case makes the strongest demands on useful tasks, reliability and compute.

By 2032, 95% of knowledge workers use agents. Each runs sixteen human-directed agents at 1.8 hours a day and forty-five autonomous agents. Of the latter, 70% work 23 hours a day; the rest average 5.8. Parallel work produces about 830 hours per user per day and an annual rate of about 340 trillion hours, on the capacity assumptions.

The share of hours using each provider rises about a fifth from 2025. After each tenfold rise in volume, prices retain 50% of their earlier level at Cloudflare, MongoDB and Snowflake, and 55% at Datadog. Existing-business growth rises three percentage points a year from 2027 through 2029, then slows to zero by 2032. Agent revenue supplies more of the later growth. Combined annual growth is 38.3%.

Limited repeat use, costly review or insufficient compute would reduce the hours delivered. Strong activity with little paid use would weaken the provider-revenue case even if the hours arrived.

Hours assumptions

Share of knowledge workers who use agents

People who actively use agents. The dots show each scenario’s 2032 setting.

The 12% starting point in 2026 is our global estimate. By 2032, the scenarios reach 70% and 95%. Surveys of businesses, employees and paid licenses provide context; none directly measures global active agent use.

Human-directed agents per user

Agents a person directs for coding, research, drafting and other tasks.

Sixteen agents per user is a scenario assumption, not an observed average.

Hours each of those agents works a day

Time actively working, averaged over a calendar day.

In our model, 16 agents at 1.8 hours each give 28.8 machine hours per user per calendar day, or 40.3 per working day on a five-day week.

Autonomous agents per user

Agents allowed to start work under standing instructions.

The Machine Economy assumes 45 autonomous agents per user. Company plans help frame the possibility but do not measure adoption.

Share working through most of the day

The rest run only when something triggers them.

This group must be actively working for the stated hours. Being available under standing instructions does not count. The Machine Economy’s 70% share is a strong assumption.

Daily hours for continuously working agents

Active hours only. Waiting is excluded.

The Machine Economy assumes 23 active hours per agent per day. Parallel agents contribute their own hours. This is agent activity, not hardware utilization, and useful results must justify the full cost.

Daily hours for event-triggered agents

Active hours added up across triggers in a calendar day.

We add up trigger frequency, duration and retries across a calendar day.

What a machine hour is

A machine hour is one hour of active execution by one agent process, with waiting and sleep left out; a child process a parent spawns adds its own hours, so a hundred agents working the same hour are a hundred machine hours. Failed work counts as demand, not as useful output.

An hour of machine work is not a fixed amount of GPU time. The capacity check in Appendix A3 is illustrative: it uses assumptions about hardware growth, workload mix and efficiency, rather than measured GPU time for every agent hour.

How the paths are built

The published scenarios retain their yearly paths. Moving a slider changes the 2032 setting and adjusts the years between it and the fixed starting point. A setting beyond the two published scenarios extends the nearer path.

Human working time uses an eight-hour day and five-day week: 5.71 hours per calendar day. This keeps the daily comparison on the same basis as machine hours.

Every year is an estimate

Every year in the hours model, 2024 to 2032, is an estimate; 2024 and 2025 are fixed and never move, and the charts begin at 2025. The 2026 starting points are modeled starting points set against the readings listed under the hours build, not a measurement of global agent activity.

The output depends on the inputs in combination, not on any one number.

The revenue model’s limits

The illustrative bill applies public list prices checked on September 29, 2026 to a stated workload. The workload assumptions draw on Claude Code studies, a LangGraph example and Cloudflare’s cost comparison. Customer discounts and included usage can make actual bills different.

The provider shares are estimates. At the starting point, roughly a tenth of hours use Cloudflare, a few percent use MongoDB or Datadog, and a few in ten thousand use Snowflake. One hour can use several providers. We need production observations to test those shares.

An ambition is not a calibration

A corporate agent-count ambition describes one company’s plans. It does not validate global adoption, utilization or the economics of recurring work.

A global adoption assumption needs a representative population, a measurement period and one definition of active use.

What we need to measure next

The workflow panel brings time, tokens and outcomes together from coding, service and research runs, including public run logs from two other coding operators, All Hands AI and METR. We still need broader coverage, direct review costs, actual infrastructure bills and repeated paid use.

A matched set of customer workflows and invoices would let us test how agent activity reaches reported provider revenue.

Research commentary, not investment advice. Scenario outputs are Skycatcher estimates on stated assumptions, not company guidance. Reported data retain their sources and dates. The four companies have not reviewed these scenarios. A case you set with the sliders is your own input, not a Skycatcher estimate, forecast or recommendation, and it is not stored. Full disclosures in Appendix D3.