Skip to content
Omkar Khadamkar

Work in progress · updated 29 Oct 2026 · notes, not conclusions

Back to Exploring
Exploring

Letting an agent act alone

A security lead sets how far an AI agent may act on its own, per kind of action and from its track record, and no step leaves a person out.

Permission, one action at a time
  1. You actthe agent only recommends
  2. You approve each
  3. Runs unless you stop it
  4. You review afterevery action still lands in a person's review list

If its accuracy on that action drops, it steps back down and says why. Set separately for each action: isolate a laptop, block a site, reset a password.

When should a security lead let an AI agent act without asking first, and how do they take that permission back?

For exampleThe agent has cut 200 infected laptops off the network correctly this quarter. The 201st is a domain controller. Same action, same track record. Should it act alone?
Update · 29 Oct 2026This ladder is now a proposal in Vela, my open design system, and it changes three things. The rungs are named by what the person does: You act (the agent only recommends) · You approve each · Runs unless you stop it · You review after. "Ask if unsure" is a policy condition, not a rung: an agent's own doubt should not decide whether a person sees a change, so low confidence can only move a change down the ladder, never up. "Act alone" is not offered: on the highest rung, every action still lands in a person's review list. Everything else on this page carries over. Read the proposal (specced, not built).

Two dials, not one

How much an agent may do alone is one dial. What it can reach is the other. The same action, isolating a laptop, can run alone on an ordinary laptop and still ask first on a domain controller.

Why it matters

Agents are fast, and asking a person every time makes them slow again. But a wrong automatic action stops someone's work, or worse. Today that trade-off is usually a hidden setting, not a decision anyone can see or review.

What I'm trying

What would show it works

Security leads who use the screen set permissions they don't later reverse, and can explain each one in a sentence.

Next step

Sketches of the track-record screen, then three conversations with SOC leads.

Related work

Nothing here is for sale until a page proves it.Next: When another AI asks your product