AI Founder Weekly

Hiten Shah puts a lead agent in charge of his coding agents

A five step setup for running parallel coding agents without drowning in chat windows.

Hiten Shah says his team has been running a lead agent on every project since February. That agent holds the plan, remembers decisions, delegates work and knows what is stuck. It can start other agents for specific tasks, which work in parallel, bring questions back to the lead and report in.

Hiten Shah
@hnshah
X
I spend far more time looking at actual work now. The agents keep track of the rest.
Nov 8, 2026 · View on X

The part people are copying is step four. Everything that needs Shah's judgment goes into one queue, where he points at what needs changing, gives feedback or approves. His response goes straight back to the agent. Feedback stays open until it is addressed, changes get tested before they land, merges get a second opinion and anything risky needs approval.

He also makes agents prove what they did. Screenshots, live pages, different options and what changed since the last round.

Why the queue matters

If you have run more than two agents at once you know the failure mode. You end up with a dozen chat windows, each one half finished, and no sense of which is waiting on you. Shah's setup moves the tracking job onto a machine and leaves the founder doing the one thing that does not delegate, which is looking at the work and saying yes or no.

Brian Armstrong described a near identical system in the same week. One lead agent per area of focus, running all the time, plus a chief of staff agent overseeing them, with sub-agents that write code, test it, chase reviews and coordinate with team members. His queue is a simple local web app, each product review comes with screenshots and a video demo, and he clears it in batches by voice and keyboard shortcuts.

Brian Armstrong
@brian_armstrong
X
It is eliminating tasks, not human judgment.
Nov 7, 2026 · View on X

Armstrong says working from one queue instead of switching between dozens of chats has cut a lot of mental overhead, and that the system is learning his preferences so first drafts need fewer corrections. He is clear that every change still goes through the same human review and security checks.

Two operators, two independent descriptions, one shape. Neither has published throughput numbers, so what you have here is a working pattern rather than a measured result. The useful read is where the bottleneck has moved. It is not writing prompts anymore and it is not writing code. It is reviewing output, and the setup both describe is built entirely around making that review fast.

Steal the numbers, not the vibes

One email a day. Under three minutes. Real MRR, real pricing, real launches from founders building with AI.

Get the next one by email

Founders building with AI, every day. Real revenue, real pricing, real launches and what they would redo. No hype sludge. Every number links out.