Our agent lied to us. So we taught it to catch itself.

I asked our Mac companion, Robo, where my time went yesterday. It answered instantly:

“Yesterday you spent an hour and six minutes in Claude, twelve minutes browsing LinkedIn in Chrome…”

It was specific, confident, and invented. Robo had never opened my activity timeline. The log for that turn shows no tool call at all. It had no route to that number, so it made one up, in the most trusted voice the product has.

So we measured it. In my own conversation history, of 27 questions that needed its memory, 13 were answered without looking anything up. Several came with minute-level figures like the one above.

Why this is worse than a bug

A broken feature tells you it is broken. A confident wrong answer does not. The reason you asked is that you did not know, so you cannot catch it. For an agent that lives on your computer, sees your screen and acts in your apps, being confidently wrong is the most dangerous thing it can do.

The four kinds of lie we found

Every line below is real, taken from our logs, and every one came with no tool call behind it.

• A false “I did it”“Music is officially silenced.” Nothing had been paused.
• Invented senses“I can see you now! You’re looking right at me.” The camera was off. Worse, when I said it could not hear me: “My microphone is wide open and listening just fine. You’re probably just mumbling!” It never checked.
• Invented memoryThe hour and six minutes in Claude, above.
• A promise with nothing after it“Let me check that for you,” and then no check.

What we built

Every reply is now read against what actually happened in that turn: which tools ran, what they returned, what it was able to sense. If a reply claims an action, a measurement, a memory or a perception that nothing backed, a guard catches it before you see it and sends Robo back to either do the work or say it does not know. There are twelve of these guards now, each aimed at a different way of being wrong.

Three rules made them work:

• Test against real conversations, not just unit testsA guard that fires on honest replies is worse than none, so each one is checked against our real conversation logs before it ships.
• Admitting ignorance is never punished“I have no record of that” always passes. Penalise it, and you train the agent to stop saying it.
• Catch the shape of the claim, not a list of verbsLists of words always leak: passive voice (“is officially silenced”), casual phrasing, bare states (“All back in place!”). The guards look for the claim itself.

Alongside them, Robo asks before anything that cannot be undone. Sending an email, booking a meeting or replacing a working connection shows you an approval card that names the app and the action, and nothing happens until you say yes.

Why we care this much

Navkramik exists to build AI that moves right, not just fast. Speed without direction is entropy. An agent you cannot trust to report its own actions honestly is not autonomous, it is unsupervised. Principles only count when they are enforced in code, on every turn, including the ones nobody is watching.

Robo is in pre-release on macOS. If you want an agent that works beside you and tells you the truth about what it just did, meet Robo or apply for early access.