A run goes well
A distillation pass reads it — the goal, the sequence that succeeded, the outcome — and writes at most one generic, reusable procedure out of it.
Auto-learning
Agents that get better at your business.
A new employee is slow in week one and fast in month three — because they learn how you work. Unyo’s agents do the same: when a run goes well, the method that worked is distilled into a skill. You approve it, and from then on it’s how that agent works.
Nothing activates without youMethod, never permissionYour runs only
Blake distilled a skill Candidate — invisible until you decide candidate Weekly report method — pull the numbers first, compare to last month, then write the two-line summary before the charts.
Trust 0.2
Skill active Blake applies it from his next run. Keep or Archive — that’s the whole interface, and that’s the point.
How a skill is born.
Three steps, and the last one is always yours.
A card arrives
The skill appears in your chat as a candidate, in plain language, with a notification so you know it’s waiting. Candidates are invisible to every agent.
You decide
Keep it and the agent applies it from its next run. Archive it and it’s gone. There is no third path — no skill ever self-activates.
The rule
Learning is approval-gated.
Agents don’t quietly rewire themselves on your account. Everything learned starts invisible and stays invisible until you press Keep — because an AI that modifies its own behaviour unsupervised is exactly the thing a business can’t audit. We built the opposite.
Born invisible You hold the pen Archived = gone Skills keep score
Narrow on purpose.
A skill is a procedure — the how. Facts about your business belong to the Neural Core; your corrections shape the memory, your successful runs become skills, and the two compound.
It only reads scrubbed material
Before anything reaches the model, emails, links and long numbers become placeholders — and it sees the names of the tools the agent used, never their contents.
It refuses to learn grudges
“This integration is broken” is not a procedure and is filtered out. A bad Tuesday can’t become a permanent belief.
It doesn’t hoard
One skill per run at most, duplicates dropped, and never more than five candidates waiting for you. Approval is a moment, not a chore.
Questions
Can a skill change what an agent is allowed to do?
No. Skills are method, not permission — they run inside the same hard toolbox boundaries as everything else. A learned skill cannot give Ashley your inbox; nothing can.
How do I know when an agent has learned something?
A notification, and a card in that agent’s chat describing the proposed skill in plain language. Nothing activates until you press Keep on that card.
What if I approve a skill and regret it?
Ask the agent to drop it — archived skills stop applying immediately. And a skill that keeps failing to help is flagged by its own score rather than lingering silently.
Does another user’s data ever leak into my agents’ skills?
No. Skills are distilled per account, from your runs only, under the same row-level security as everything else you own.
Hire a team that learns the job.
The tenth week should not look like the first. 2,500 credits free — no card.