Your agent works. It's still stuck in pilot.
Most teams we talk to are in the same spot. The agent does well on the cases they tested, the evals look fine, and it still isn't allowed near real customers unless someone is watching it.
When we ask why, we usually get some version of the same question. What does it do at 2am when a customer asks about an exception that was agreed in a meeting and never written down?
Right now it answers anyway, and it sounds sure of itself. Everyone on the team knows that one made-up refund policy is enough to put the project back on the shelf for a quarter.
The usual fixes are a bigger model, a longer prompt, or a human approving every run. A bigger model still doesn't know what your ops lead decided on Tuesday. A longer prompt only holds what somebody remembered to write down. Approving every run works for a couple of weeks, until the person doing it has four hundred items a day and starts clicking through them without reading.
All of these try to make the agent know more, and we think that's the wrong goal. Nobody at your company knows everything either. A new support hire doesn't know your return policy on day one, and that's fine, because they can ask the person sitting next to them.
The agent has nobody to ask.
So that's what we built. Fruxon gives the agent a list of real people, what each of them is responsible for, who it's allowed to ask and in what order, which channel to use, and what to do if nobody replies. At 2am it messages Sarah in Ops on Slack and waits for her. The customer doesn't get an answer until she's given one.
Over time it needs to ask less, because last month's questions become this month's answers. It never stops asking completely, and we don't think it should. Senior people at your company still check with legal before they sign things.
We didn't come up with this from scratch. We were engineers at Siemplify, which made security operations software, and stayed on at Google Cloud Security after Google bought the company in 2022. Security automation has worked this way for years. It runs the lookups and containment steps it's approved for, and when something needs a human decision it pages a specific analyst and waits. In security, that's the basic requirement for letting automation run at all.
When we started building with LLM agents, that part was missing. That's why we started Fruxon.
Some things should always go to a person: payment reversals, apologies to customers, anything you can't undo. Fruxon lets you decide where that line sits, and move it later once the agent has earned it.
The org chart doesn't go away. The supervision does.
What we want is pretty simple. At some point you look back at a few months of the agent asking when it should have and getting the rest right, and you give it more to do. We hope that's a boring decision by then. It means the agent is out of pilot.
Co-founder & CEO
Co-founder & CTO