AI · Idee
Agents need boundaries, not chains
Autonomy is not a switch but a radius. What an agent may do comes down to three questions: What can it break, who notices, and how do you get back?
In four sentences
- The question is never how autonomous an agent can be, but how big the damage is when it fails.
- An agent can take reversible steps on its own. Irreversible ones belong in front of human eyes.
- Without a log there is no trust: if you cannot read back what happened, you cannot hand over more responsibility either.
- Boundaries don't make agents weaker — they make them deployable.
The radius counts, not the capability
Most discussions about agents revolve around capability: what the model can do, how many steps it sustains, which tools it operates. In production, that is the second question. The first is: What happens when it goes wrong — and who gets hit?
An agent that pre-sorts invoices has a small radius. One that pays invoices has a large one. Same model, same capability, two completely different systems. Determine the radius first, and you can be generous with autonomy afterwards.
„An agent with no way back is not a tool — it is a risk with good marketing."
Reversible first
The most useful sorting I know is banal: Which steps can be undone, and which cannot? Writing a draft, kicking off research, enriching a record — all reversible, all agent territory. Moving money, canceling contracts, deleting customer data, sending out emails — irreversible, so with a hand on the lever.
This is not distrust of the machine; it is normal system design. We build confirmation dialogs before deletion for humans, too. The only difference is that an agent errs a thousand times faster.
What you need to see
Three things belong in the build from day one, or the agent will never grow up: a readable log of every step, a hard limit on costs and retries, and a way to abort running operations. Sounds unspectacular, but it decides whether after three months someone trusts the system with more — or quietly switches it off.
Trust in agents grows the same way trust in new employees does: through manageable tasks, visible results, and the experience that mistakes surface early. Except that here, onboarding is an architecture decision.
Questions about this
How much autonomy should an AI agent get?
As much as the blast radius allows. It can take reversible steps on its own; irreversible ones need human approval — regardless of how capable the model is.
What does a responsible agent need technically?
A readable log of every step, hard limits on costs and retries, defined failure paths, and the ability to abort running operations.
Don't guardrails make agents useless?
Quite the opposite: only clear boundaries make it possible to give an agent real tasks in the first place. Without them, it stays stuck in pilot mode.
