The things people actually ask.

Your agents do not keep office hours. Neither does our monitoring.

Every agent is watched automatically, around the clock. If a run fails, stalls, or starts behaving oddly at two in the morning, the system catches it and one of us is paged — we do not wait for you to notice on Monday.

Real hours, honestly stated.

People are at their desks 8am to 6pm Eastern, Monday to Friday. Outside that we are reachable for anything genuinely broken, and we answer — but we are not going to claim a night shift we do not staff. What runs overnight is the monitoring, and that never sleeps.

You find out from us first.

The failure you would normally discover days later, when a report did not arrive or a follow-up never sent, is the one we have already caught and usually already fixed. That is the part worth paying for.

One run is one agent completing one job once — a weekly report written, an inbox triaged, a batch of expenses categorised.

We meter runs rather than agents because cost tracks how often something runs and how much it has to read, not how many agents exist. Two businesses on the same tier can otherwise differ tenfold in what they actually consume.

Nothing stops working and nothing is throttled mid-job.

Additional runs are billed at cost on your next invoice, itemised, and we will tell you before a pattern of overage becomes a surprise. If you consistently exceed a tier, we move you up rather than quietly billing overage forever.

We evaluate it against your actual agents, on your actual work, before anything changes. Not against a benchmark.

If it is genuinely better for a job, we move that job onto it. If it is better at some things and worse at others, we move only the parts that improve. Different agents can run on different models, because the best model for reading a plan set is not necessarily the best one for writing a follow-up email.

This is the real risk, and it is why we test rather than auto-upgrade.

We have seen a newer, higher-benchmarking model fail eight times out of eight on a production job that the previous model had completed successfully for twenty consecutive runs. Same task, same instructions, silently produced nothing. A model that looks better on paper can be worse at your specific work.

So nothing switches until it has produced correct output on your jobs, and every agent can be rolled back to the model it was running on before.

Your tier price does not change because we changed a model. You are billed on agent runs, not on tokens, so what a run costs us behind the scenes is our problem to manage rather than a line item that moves on your invoice.

When a model gets cheaper — which happens regularly — that shows up as more headroom in your run allowance rather than a price rise on our side.

No downtime. Your agents keep running on their current model the entire time a new one is being evaluated. The switch itself is a configuration change, not a rebuild — it takes effect on the next scheduled run.

Evaluation takes as long as it takes to see the new model handle a real cycle of your work. For an agent that runs daily that is days; for one that runs monthly it is longer. We would rather be slow and certain than fast and have you find the regression.

No. Agents are built against the job, not against one vendor's model, and we route each one to whichever model does that job best at the lowest sensible cost.

That matters commercially as well as technically: if a provider raises prices, degrades a model, or has an outage, we move the work rather than passing you the problem.

The question we ask before you sign anything.

How much new revenue would this have to produce to pay for itself? Take the first-year total, divide by your gross margin, and that is the number. If it is not plausible for your business, we tell you to buy a smaller piece.