Agent unit economics: making the per-task math survive contact with reality
Agents multiply model calls per user action by 10–50x. Without aggressive reuse, the unit economics of agentic products simply don't close.
Chat products spend one model call per user message; agents spend a planning call, several tool-selection calls, retries, reflections, and a synthesis — ten to fifty calls per task, each billed. Multiply by your task volume and many agentic business models are underwater before the demo ends. The category's dirty secret is that the unit economics rarely close at list prices.
Crowkis attacks the multiplier at every joint: semantic caching collapses the fleet's repeated questions, reasoning reuse recycles planning skeletons (the priciest tokens agents emit), the tool-call cache replays deterministic tool results, and the conversation cache stops multi-turn context from being re-purchased turn by turn. The 10–50x multiplier deflates toward the irreducible novel core of each task.
Five agents asking one question should cost one answer.
Per-key budgets make the economics governable per agent: each gets a virtual key with hard walls, so the cost of any agent is bounded by configuration rather than by hope. Finance can finally price a task with confidence intervals instead of incident reports.
The bottom line
Teams shipping agents at sustainable margins all converge on the same architecture: a memory layer between the fleet and the providers. That layer is a product now. It deploys in five minutes.