You pay for the software doing the work, the same way you'd pay for a seat if a person were doing it. Console access is unlimited and always has been — the whole point is that three teams read the same record.
Count only the agents you'd let a rule act on. Traced-only agents are free and unlimited, so most teams land far lower than they expect.
Actions pool across your whole estate, so a busy agent borrows from a quiet one rather than triggering an overage. Beyond the pool it's $0.60 per 10k actions, and we'll tell you before you get there, not after.
Every one of these is billable somewhere in this market. Charging for them would undercut the reason the product exists.
Including the one about whether this model quietly punishes you for building well.
Yes, and the agent count is unlimited while you do. You'd have full traces, replay and cost accounting across the whole estate for nothing.
What you wouldn't have is anything that stops an action, any risk scoring, or alerts reaching your on-call. Teams usually run that way for a quarter, then find one thing they wish had been blocked.
Nothing disappears. Full step-level replay covers 30 days; after that you keep 12 months of metadata — what ran, when, which tools, which systems, what it cost.
So you can still find the day an agent touched something it shouldn't have. Opening that run and watching it back is what needs a paid plan. We'd rather you hit a locked door than an empty shelf.
A fair-use ceiling rather than a wall. Past it we sample the repetitive traffic — repeated reads, unchanged queries — and never sample anything above baseline or touching a policy-relevant path.
A supervision tool that stops recording mid-incident is worthless, so the ceiling degrades rather than cuts off, and it starts a conversation instead of a bill.
A distinct identity, holding its own credentials, with at least one enforcement rule pointed at it. Ten instances of the same agent behind a queue is one agent. Two agents sharing a codebase but holding different access is two.
Traced-only agents never count, however many there are. We count what your security team would count, because they're the ones who have to review it.
It would, and that's the honest weakness of the model. Cramming six jobs into one agent to save $59 a month is a bad architecture decision that our pricing would have caused.
So two things. Agents under 10,000 actions a month don't count toward your total, which covers most of the small specialised ones. And past 100 agents the per-agent rate falls faster than the count rises. If you still find the model bending your design, tell us — that's a pricing bug on our side.
Enforcement stops. Your agents keep running — we won't hold your production hostage over an invoice. You get 90 days of full console access, then a final export of the complete trail.
The hash chain stays independently verifiable afterwards, with or without us.
Observe costs nothing and stays that way. If you never move past it, you'll still know more about your agents than you do today.