Concepts
Agents & autonomy
People, agents, and guardrails.
3 min read
Agents and autonomy
Every agent in AIBOS is given a defined role the moment it's created — never a blank slate that figures out its own job as it goes. That single fact is what makes it possible to know what an agent will and won't do before it ever does anything.
Three kinds of agent
| Kind | What it actually does | What that feels like |
|---|---|---|
| The one you talk to | A thinking partner inside a live session — research, drafting, working through a problem together | You see its reasoning as it happens, and it only runs while you're there with it |
| The one that works around the clock | Runs a specific, defined job continuously, without you watching | It reports what it finished, not its whole train of thought — and it operates inside a set budget and mandate, so it can't just decide to do more |
| The one that handles email and calendar | Always on, representing you to other people outside your sphere | Anything that genuinely needs a human gets handed straight back to you |
Worth being honest about: the email-and-calendar kind is both the newest of the three and the hardest to get right, precisely because it's autonomous and talking to the outside world at the same time — exactly where a mistake is most visible and costs the most. The safety rules for it were built and shipped before the agent itself was considered mature.
The caretaker
One more role sits apart from all three of these: the caretaker, a nightly job rather than an agent you'd ever hold a conversation with. Every sphere gets exactly one, running automatically overnight on the schedule set up during installation. It tidies loose ends, checks the sphere against the standard it's supposed to match, and leaves a short report. It never decides anything on your behalf — its whole job is maintenance, nothing more.
Your lead agent
The first agent any sphere receives is your lead agent (the standard calls it the HeadOfSphere), created the moment the sphere itself is. The standard describes it plainly: your co-creator and thinking partner — the first agent you get, and for as long as you want it that way, the only one you strictly need. It moves between the three modes above depending on what you're asking of it at the time.
What it just does on its own, versus what it checks with you about first, comes down to one simple distinction: small, reversible things inside its own job, it just does. Anything bigger — something hard to undo, something that reaches outside its own mandate, something genuinely consequential — it surfaces to you before acting, rather than after.
Born with a defined role, never drifting into one
Every agent is typed the moment it comes into existence, through the same process that creates it in the first place. The role assigned at birth is the role an agent keeps, until someone deliberately changes it.
A prompt is input, not an order
One idea governs how an agent is meant to treat what you ask it to do: what you type is information it weighs against what it already knows about the rules of the sphere it's working in, not a command it executes blindly regardless of what it notices along the way. An agent that does exactly what it's told, every time, without ever raising a concern, is doing less than its job actually requires. Part of its job is to push back — to flag a real risk, a likely mistake, or a request that runs against something it already knows about how the sphere is supposed to work.
That pushback is part of what an agent is actually for. It's judgment applied on your behalf, not resistance to what you want. An agent that never disagrees with anything is either only ever handed trivially safe work, or it isn't thinking hard enough about the work it has.