8 min read

The real cost of running AI agents

Licence fees are the small number. Token spend, retries and human time are the real bill — and the reason uncapped agent systems produce surprise invoices.

Last updated 18 August 2026

Every pricing page in this category shows you the smallest of the three numbers you will actually pay.

The three lines on the real invoice

1. Licence or subscription

The published number. For open-source frameworks it is zero, which is exactly why they appear cheapest and frequently are not.

2. Model API spend

Usually larger than the licence, and the line that behaves badly. It scales with how chatty your system is, not with how much value it produced. A verbose multi-agent conversation and a tight single-agent task can differ by an order of magnitude for identical output.

3. Human time

Always the largest, never on any pricing page. Setup, debugging, prompt tuning, integration maintenance, and the meeting where you work out why it stopped. At any realistic hourly rate this dwarfs the other two in month one and often keeps dwarfing them.

Why token spend surprises people

Three mechanics do most of the damage, and none are obvious when you start:

  • Retries. A failed step is not free. Systems that retry aggressively can triple the cost of a task that eventually succeeds.
  • Conversation overhead. Agents that negotiate re-send context on every turn. The context is billed every time it is sent, not once.
  • Loops. Two agents that cannot agree will keep going. Without a hard stop this runs until something external intervenes — usually the invoice.
The structural fix

A hard per-agent monthly ceiling turns every one of these from a financial event into a paused agent and an alert. It is unglamorous and it is the single most valuable cost control in agentic systems. If the tool you are evaluating cannot do it natively, you will be building it.

A fair way to compare options

Build the comparison on total cost of ownership over twelve months, not the sticker:

  • Licence or subscription for the year
  • Estimated model spend at your realistic volume — then double it, because first estimates are always low
  • Setup hours multiplied by a real hourly rate
  • Ongoing maintenance hours per month, multiplied by twelve
  • The cost of the thing not getting done during the weeks it is broken

That last line is the one people leave out, and it is often the largest. A framework that takes six weeks to get right has six weeks of unposted content in it.

How we price against that

We fold model costs into the plan price and cap every agent, so lines one and two are a single predictable number and line three is close to zero for you — it is our time, not yours. That is the entire reason a managed service can be worth more per month than a free framework.

The plans are published: $499, $899 or $1,499 a month by how many departments you automate, plus a one-time $750 build. The full framework-versus-service argument is in this guide.

Keep reading

18 Aug 20268 min read

Why multi-agent AI systems fail in production

Most multi-agent demos work and most multi-agent deployments do not. The reason is error compounding — and the fixes are structural, not model upgrades.

Read
Want this done rather than explained?

We build the agent team, connect it to your accounts and supervise the output. Twenty-minute call · See pricing