The short answer
Most first agents we put in production cost between $45k and $110k to build and $1k to $4k a month to run. The range depends less on the model and more on how many systems the agent touches.
What you pay for
About 60% of the budget goes to integrations and permissions, 25% to evaluation and guardrails, and only 15% to prompts and model work.
const run = await agent.execute({
task: "follow_up_overdue_invoices",
requireApproval: ["send_email"],
})
Rule of thumb: budget one week of evaluation work for every system the agent can write to.
Model costs
For a finance agent handling 2,000 runs a month, model usage came to $380. Logging and monitoring cost more than the model itself.
Prices keep falling, so we design every agent to switch providers without a rewrite. That keeps the model line small and predictable.
Evaluation and guardrails
Every agent we ship has a test set built from real cases: the invoices, tickets or documents it will actually see. We run it on every change and report accuracy weekly. Anything that sends, pays or deletes waits for a person by default.
Support after launch
Plan for a small monthly retainer. Models change, your systems change, and edge cases show up in production. Most of our agent clients spend $1k to $4k a month on monitoring, tuning and new skills.
Our numbers
Across our last ten agent projects the median build was $68k over six weeks, and the median monthly run cost was $2.1k including models, hosting and support.
