How it stays honest
Giving real autonomy to something that fabricates results would be reckless. So honesty isn't a nice-to-have in Mylo — it's the mechanism that makes autonomy safe to grant. An agent earns the right to run on its own by proving it tells the truth about its own work.
The danger with autonomous AI
Language models are fluent — including when they're wrong. A model will happily say "done!" for a task it didn't finish, cite a source that doesn't say what it claims, or report success it can't actually demonstrate. That's a harmless quirk in a chatbot you're watching. It's a serious problem in an agent acting on your accounts while you're away.
The fix isn't a better-behaved model. It's a verifier the model doesn't control.
It verifies its own work
Before a Mylo agent claims a task is complete, a separate verification step checks the claim against real evidence — the actual result, the actual source, the actual state of your world. The agent can't simply assert success; it has to show it.
The rule is absolute: an agent never marks something done that it can't back with evidence. If it can't verify, it says so honestly — "I couldn't confirm this" — instead of fabricating a pass. No fake checkmarks.
This is why "verified" means something in Mylo. A green result isn't the model's opinion that it did well; it's a check that survived an attempt to disprove it.
When it can't be sure, it abstains
Some things can't be deterministically verified — "is this a good strategy?", "is this the right tone?". For those, the agent doesn't bluff a verdict. It abstains honestly: it surfaces the evidence, says it can't be sure, and routes the judgment to you. An honest "you decide" beats a confident wrong answer every time.
Unbreakable rules
Above every agent sits a small set of safety rules the OS enforces no matter what an agent is asked. They're not instructions an agent can be talked out of — they're invariants of the system itself: what may touch your secrets, what requires your sign-off, what can never be done irreversibly without you. This is what makes it safe to let an agent's autonomy grow.
Trust is earned, then extended
Because an agent proves its honesty continuously, trust isn't a switch — it's a track record. An agent that reliably backs its work earns room to run further without you checking every step. One that can't stays close, proposing and confirming. (More in Autonomy & proactiveness.)
Next
- —It gets better as you use it — how a trustworthy agent compounds.
- —How an agent runs — where verification sits in the loop.