RRochak Shrestha

Build log · · Rochak Shrestha

An approvals desk for AI agents that do real work

My fifteen minutes a day used to go on finding what the agents wanted from me. Then every loop started filing on one page.

Early on, my fifteen minutes with Kata Jaam began with a search. Each agent reported into its own file, its own channel, its own log. I opened one, then the next, then the next, and by the time I had found everything the agents wanted from me, most of the fifteen minutes had gone on finding it. I was finding the work instead of judging it.

Human in the loop for AI agents works when the loop is one page, not a chat window. Today every outward action my agents propose files as a row on a single approvals desk and waits for my click. That desk, a dumb CLI underneath it and sending caps enforced in code are why Kata Jaam, an event-discovery product I run alone, needs about fifteen minutes of my attention a day (as of 17 Oct 2026).

Three places to look

Kata Jaam runs on a rota of Claude Code agents with written briefs: inboxes, event scouting, catalogue QA, social drafts, SEO, retention, organizer success. Each one reads, judges and drafts. None of them is allowed to act on the world alone, so all of them end in the same place: something waiting for me.

The fifteen minutes are the whole budget. I run the company alone, next to a full-time job. When every loop leaves its output somewhere different, the walk between them eats the budget before a single decision gets made.

One page for everything that leaves

So I put every loop on one desk.

If an action leaves the building, it is a row: a DM to an organizer, an email, a social post, the approval of a listing, a paid placement, a delete. No agent does any of these on its own. The agents read, judge, draft and file. I click. All of them file to the same page, and the search at the start of the morning stopped.

Sorted by what is already wrong

Once everything was in one place, the next question was what to look at first. Rows sort into four groups, and the order is most of the design:

  • Failed after approve: I said yes and something broke. This is first because it is the only group where the world is already wrong.
  • Overdue: a reply or a task that has waited past its window.
  • Decisions: drafts that need a yes, an edit or a no.
  • Approved and needs a hand: things I said yes to that still need a human step.

There are keyboard shortcuts for approve, edit and skip, so clearing a row takes one keystroke. The badge counts only what needs me now. A run log sits beside the desk and shows what each loop did and what it cost, so when a row looks odd I can see the run that produced it.

A yes runs at once

An approved row executes immediately. There are no "send later" conditions inside drafts. A condition I cannot see is an action I did not approve. If an agent wants something sent on Friday, it files the row on Friday.

One path to production

The desk is only worth trusting if a click means exactly what the row says. Under it, every write to production goes through one thin CLI that calls the admin's existing API. The CLI has no AI in it. Every verb has a dry run, a likely duplicate is refused, and it logs what it did. The agent decides the fields and the CLI writes them. There is exactly one path to production and I can read it.

The hard limits live in code too. Daily sending caps are enforced by the system, and each channel has its own daily cap. A prompt is only advice to a model, so a limit that matters cannot live there.

The two other ways to keep a human in

There are two obvious alternatives to a desk, and I wanted neither.

Full autonomy means the agent acts and you read the log afterwards. It looks efficient until the first wrong message reaches a real person. Then you are cleaning up in public, and the only fix available is to turn the agent off.

Chat-style confirmation means the agent stops and asks "shall I send this?" at every step, inside a conversation. It keeps a human in the loop, but it puts the human in the agent's loop, on the agent's schedule. You have to be present while it works, and the questions arrive one at a time with no ordering.

The desk sits between the two. The agents work when they work, and I decide in batches, sorted by urgency, a keystroke per row. How much a loop may do alone is a per-loop setting, and the approval rate decides it.

The same tap inside Livio

The AI coach I built inside Livio is the same discipline pointed at a user instead of a founder. It calls thirty-plus tools to read and change a member's real health data, and every change it proposes is confirmed by the member with one tap. Numbers are computed in code, never by the model. The case study has the details.

What still gets my click

The desk catches what I can see. It does not catch a judgement error that looks fine. A polite, well-formatted draft built on a wrong reading of the situation will get my click, because it looks right in the row. The usual shape is a reply that answers the question the organizer asked when they meant a different one. The weekly retro loop reads the journal and proposes diffs to the briefs, which is how those errors become rules, but only after they happen.

It also makes me the bottleneck. Nothing outward moves until I clear the page. On a day I do not open it, the company waits. Overdue rows sort above everything except failures, so the next time I open the page the replies that have waited longest are near the top.

And it does not move the numbers that matter most. Week-four retention at Kata Jaam is under ten percent (as of 17 Oct 2026). The retention loop reads the cohorts every Thursday and drafts the weekend push plan, and I approve it.

Eight months of clicks

Every loop starts in phase one: it files drafts and sends nothing. I widen autonomy per loop, and only when the approval rate says the drafts are right.

Eight months in, outward messages still wait for a click, and that is fine. Approving a good draft costs a keystroke. Recalling a bad DM to an organizer costs more than that. The fifteen minutes now go where I wanted them to go at the start: on the rows, one keystroke at a time, from the top of the page down.

I build this kind of agent setup, approval queue first, for other businesses: AI operations.