2026-05-10 · 12 pages
Most workflow tools price per "task" or per "operation." This was reasonable when each task was a deterministic move-data-from-A-to-B step costing a fraction of a cent. Agents broke that model. A single agent run can spend pennies (one tool call, one short LLM response) or several dollars (a long voice session, a deep browser exploration, a planner loop). Charging a flat per-run price either bankrupts the platform or overcharges the operator.
This paper proposes a three-axis meter for agent platforms: tokens (LLM input and output), voice seconds (STT plus TTS), and browser seconds (headless or interactive browser time). Each axis maps cleanly to an upstream cost the platform actually pays, so the operator's bill is an honest pass-through plus a defined margin.
The three axes are not equivalent. Tokens are cheap and high-volume. Voice seconds are middle-cost and bursty. Browser seconds are the most expensive and the most variable: a residential proxy plus a Browserbase session plus planning tokens can hit fifteen cents per minute in the worst case. A flat per-run price hides this variance from the operator. A three-axis meter exposes it and lets the operator make informed decisions about which agents to keep running.
The workspace daily cost ceiling sits on top of the meter as the kill switch. It is the simplest possible budget primitive: a single dollar amount per UTC day, applied to the workspace. When it is hit, every running agent halts. We argue this is the only kill-switch contract operators actually trust, even when more sophisticated controls (per-agent ceilings, per-tool ceilings, adaptive throttling) are available alongside it.
The paper includes a worked example pricing model showing how the three-axis meter and the daily ceiling combine to produce predictable monthly bills for typical operator workloads, plus a discussion of edge cases (multi-day batch jobs, scheduled work that straddles the UTC reset, agent-to-agent calls that cross workspaces).