Let AI act on your business — without letting it act unsupervised
The job, as most teams state it: "we want agents doing real work, and we cannot explain to a customer, an auditor or our own board why an AI did something nobody approved." Oneop's answer is not a policy document. It is a promotion ladder built into the product: every agent starts powerless, earns write access one autonomy level at a time, and leaves a full reasoning trace at every level.
Agents run on every plan.
Autonomy as a dial, with the evidence trail to justify each notch
A demo cannot be laundered into your audit trail
Run any agent with simulated: true and you get a demo trace with no model call, zero credit usage, and a null confidence rendered as an em dash. Every run is badged Simulated or Live.
Ceilings that are numbers, not adjectives
Six reasoning iterations by default with a hard cap of twelve, a 6,000-token budget, a 45-second wall clock, and three actions per run. Hit any of them and the run ends as escalated.
The line that stays closed regardless
Finance and stock actions are fail-closed until switched on separately, regardless of autonomy level and regardless of a human approval.
The evidence trail your reviewer will ask for
Every live run stores the sequence that actually happened: each step's thought, the tool it chose, the real observation it read back, the guardrail result, the model, the token count, and the reason the run ended. Confidence is model-derived from a scoring pass or null — never fabricated.
Specific, checkable facts — not a testimonial
No accuracy percentages or run statistics to show you yet — here are mechanisms you can check yourself.
A run that hits a ceiling ends as escalated with the exit reason recorded — it does not invent an answer to finish.
Prompt-injection blocking and PII redaction run on every model call, every plan, with no setting and no administrator able to disable them.
Every call passes one metered choke point — a build-breaking CI test fails if any agent code path bypasses it.
The modules that carry the dial
Agents run on every plan, including Free
Free includes 50 AI credits a month — enough to run real suggest-mode traces and decide whether the promotion ladder suits you. On Business, bring your own Google AI key and run uncapped.
| Plan | Monthly | AI credits / month |
|---|---|---|
| Free | $0 | 50 |
| Starter | $19 | 1,000 |
| Pro | $59 | 5,000 |
| Business | $129 | 15,000 |
Questions people ask first
What exactly happens when an agent wants to write something?
It does not execute the write. Every proposed action routes through one governed dispatch function that applies the agent's autonomy level: suggest logs a notification, approve queues a real human approval, autonomous dispatches. There are 18 action types across CRM, tickets, tasks, email and drafts, and the finance and stock subset is additionally fail-closed until you switch it on separately.
Can we cap what a single run can consume?
The caps exist whether you set them or not: 6 iterations (hard cap 12), 6,000 tokens, 45 seconds, 3 actions per run, plus a configurable action limit. Beyond the run, every call is metered against the workspace's credit balance — an agent action weighs 5 credits — and with enforcement on, the cap refuses the call rather than billing overage.
How do we know a demo run from a real one?
The product distinguishes them structurally, not cosmetically. A simulated run makes no model call, writes zero credit-usage rows, and carries a null confidence; the run history badges every row Simulated or Live. You cannot present a simulation as production work inside Oneop, and neither can we.
What does the audit record of a run contain?
The trigger, every reasoning step's thought, the tool chosen, the actual observation returned, the guardrail result, the proposed actions and their dispositions, the model, the token count, the credit charge, and the exit reason — stored on the run row at execution time, not summarised afterwards.
Is any of this available on the free plan?
Yes — the AI Workforce is on every plan, including Free, with the same guardrails, ceilings and autonomy modes. Free includes 50 AI credits a month, which is enough to run real suggest-mode traces and decide whether the promotion ladder suits you. The configurable guardrail layer on top — tool allowlists, action-value ceilings, confidence thresholds — is available on Pro and above and switched on when you want it.
What if the agent gets prompt-injected by content it reads?
Detected injection patterns block the call outright — this check is unconditional, has no setting, and runs on every plan alongside always-on PII redaction. It is a defence layer, not a guarantee; the reason the autonomy ladder and the action ceilings exist is precisely that no input filter is perfect.
Get started
Start it in suggest mode. Promote it when it earns it.
Open a free workspace, point an agent at real data, and read its full reasoning trace before you let it write anything. No credit card, no sales call.
Sign-up opens when the app launches. We will email you once — no marketing list.