Self-evolution

An agent gets better at your work the more you use it. It learns the procedures you rely on, and its scheduled jobs ride those procedures as they sharpen. One rule holds it together: nothing durable changes without telling you.

Two ways it improves

Skills it learns. When the agent works out a real procedure during a task, an API quirk or a multi-step workflow, it captures that procedure as a reusable skill and loads it in later sessions. You solve it once; the agent keeps the method.

Jobs that ride those skills. A scheduled job states a goal and leans on the agent’s skills for the steps. Improve a skill once, and every job that links it inherits the improvement on its next run, with no edit to the job.

The two share one library of skills. Sharpen a procedure in one place, and every session and every job that uses it grows sharper too.

The line it holds

The platform never rewrites your durable settings on its own. A skill changes between runs and a job picks up that change automatically, because the runtime reads the live skill on every run. A job’s saved instructions are different: they change only when the agent makes a deliberate edit, and when it does, it tells you what it simplified. You see the change, and you can ask for the original back.

Why it works this way

Software that edits its own configuration silently becomes a moving target, and you can no longer reason about what it will do next. Freezing the configuration would kill the improvement. So the platform does neither: it lets the agent improve and makes every durable change legible. Months in, you still know exactly what each job does.

That audit trail is the feature, not a logging detail. It is what lets you run an agent unattended and keep trusting it.