When does handing a task to an autonomous agent actually pay off? Drag the costs and watch the model decide — the same task-based framework behind How AI Agents Reshape Knowledge Work, made playable.
Yang, Zyskowski, Yonack & Ma (2026) · arXiv:2606.07489 · model from §3 + Appendix B
Cost structure
Assistant fixed cost fc
Assistant cost per step mc
Agent fixed cost fa
Agent cost per step ma
Budget B
Total effort/spend the worker can deploy across all tasks.
Break-even step count s*
1.06
Agent wins above this many steps
Agent access
Toggle the second world on/off
Load the paper preset to verify
The affordable frontier
unlocked by agentdone in bothdroppedunaffordable
Assistant only
Assistant + Agent
Cost per task vs budget
Where the surplus comes from
Task portfolio
Task
Steps (s)
Value (v)
Assistant
Agent
Status
How it reads. Each task takes s steps and is worth v. A mode's cost is fixed + perStep×s; the worker then solves a 0/1 knapsack — pick the highest-value set of tasks that fits the budget. Agents charge more to set up but less per step, so they only beat the assistant past the break-even s* = (fa−fc)/(mc−ma). Turning agent access on can expand the frontier (Prop. 1–2) and lift total surplus (Prop. 3). The default preset reproduces the paper's Appendix B example exactly: surplus 33 → 253.
Caveat. This is the conceptual model, not the paper's measured production data. It illustrates the forces; it does not predict your real costs. Built for an article on arXiv:2606.07489.