The human gates
Where people belong in an agentic system
"Human in the loop" is the most repeated phrase in AI right now, and it usually means a checkbox. Somebody has to approve every step, so nothing bad can happen.
I built it that way first. It doesn't work.
The bottleneck was me
After a few months of running the business with agents, I asked them to audit the flow of decisions. The median time between an agent asking me a blocking question and the answer being recorded was eight calendar days.
The slow ones were not hard decisions. They were one-word approvals on work that was already specified and already tested. "Yes, merge it." "Yes, resume it." I had to come back to them later, and later was a week.
Every one of those approvals felt like safety. Together, they were the single biggest drag on the whole system. A human who has to click on everything isn't supervising. They are a queue.
Gates, not checkboxes
The fix was to stop asking where a human should approve, and start asking where a human must decide. Those are very different lists.
There are things no agent decides in our system, ever:
- Money. Any spend, price change, discount or commitment.
- Anything outward. An email, a post, a message to a customer or partner, anything published.
- New audiences or new campaigns. Who we talk to is a business decision, not an execution detail.
- Consent and legal basis. What we're allowed to do with someone's data.
- Evidence. Logs, ledgers and test records are never edited to make something pass.
Those are the gates. The human stands there, and nothing gets through without them.
Everything else can move without me, but only when five conditions hold at once. The behaviour was already decided. It can be rolled back without me. Tests pass and the end state was checked, not just the exit code. Nothing leaves the building. It's written down the same day, with the rollback path. If any of the five is uncertain, the agent asks.
That one page removed most of the queue overnight, and the system got safer, not riskier, because the rules are explicit instead of living in my mood that day.
Review doesn't have to be human
On one software project, code now ships through automated review. One agent builds, another critiques, another reads the result cold, the test suite has to pass, and every change can be reverted. No human clicks "approve" on a pull request.
People find that scary until they look at what a tired human review at 11 p.m. actually catches. What I kept for myself is the moment it goes live. The machine proposes the timing, I answer yes or not now. That's a gate. Reading diffs at midnight was a checkbox.
The shape of the ask
When a decision does reach me, the format matters as much as the rule. Agents bring me options, not open questions. Two or three choices, one marked as recommended, with the reason in a sentence. I answer with a number, from my phone, in the channel I'm already in.
An open question costs me twenty minutes of thinking. A good option list costs me ten seconds, because the thinking was done before it got to me.
Where people belong
So the answer to "where do people belong in an agentic system" is not "everywhere, just in case". It's at the gates: money, reputation, other people, the law, and the truth of the record. There, the human is not a formality. They are the reason the system can be trusted to run on its own the rest of the time.
The human gates. Everything else should flow.