Letting Agents Write, Keeping Humans Responsible

This is a working note. I’m using it in production while I write it, so expect it to change.

I’ve been letting AI agents do real work across my projects — drafting content, writing code, running analysis. The uncomfortable question that kept surfacing wasn’t “is the output good?” It usually is. It was: when something goes wrong, who is responsible?

The pattern I’ve landed on (for now)

An agent can generate anything. It cannot publish anything. Between the two sits a gate a human has to open.

Concretely, that’s a status field. Agents commit their work as draft. Nothing on the live surface reads a draft. A human reviews and deliberately promotes it. The agent proposes; the human disposes.

Two properties I care about

The gate is explicit, not implicit. It’s a flag with three states, not a feeling that something is “probably fine.” A reviewer flips it on purpose. Responsibility attaches to that action.

Failures are loud. If an agent produces malformed output, the build should break — not silently publish half a thought. Silent corruption is worse than an obvious error, because you find out later, from someone else.

What I’m still unsure about

• At what point does gating everything become the bottleneck the automation was meant to remove?
• Can the gate itself be tiered — full review for some categories, spot-checks for others — without eroding the accountability that makes it work?
• Does “a human is responsible” survive contact with scale, or is it a comfortable story I tell while volume is still low?

I don’t have clean answers yet. That’s why this one lives in progress.