The Agent OS · 1 min read

How an agent runs

A chatbot answers once and stops. An agent runs a loop — it thinks, acts, checks the result, and goes again until the goal is actually met. That loop is the difference between a tool that responds and an agent that finishes.

One turn isn't enough

Ask a chatbot to "reconcile last month's invoices" and you get a description of how you might do it. The work is still yours. That's because a chatbot runs once: prompt in, text out, done.

Real work doesn't fit in one turn. It needs a first step, a look at what happened, a correction, a next step — over and over until it's genuinely complete. So an agent doesn't run once. It loops.

The loop

Every Mylo agent runs the same cycle, powered by its six parts:

It exits the loop when the goal is met, when it hits a limit it set for itself, or when it needs a human decision — never just because it ran out of turns mid-thought.

It chooses how to think

Not every task deserves the same effort. A quick lookup shouldn't cost the same as a multi-step migration. So before it works, an agent picks a strategy:

• —
  Simple, direct tasks run straight through.
• —
  Tasks that need tool use interleave thinking and acting, step by step.
• —
  Complex work gets planned up front, then executed against the plan.

And it can change its mind mid-task — if the situation turns out different from what it expected, it adapts the approach instead of charging ahead. This is where an agent's judgment lives.

Why the loop is safe

A loop that runs on its own could be dangerous — so the Check step isn't optional. Every pass, the agent verifies its own work, and it can never mark something done that it can't back with evidence. The loop can run a long time without you; it can't run dishonestly. (More in How it stays honest.)

Next

• —
  Autonomy & proactiveness — what the loop produces when it runs on its own.
• —
  How it stays honest — the verification that makes the loop safe.