One evening with a hundred agents

I did not plan to measure my life in Cloud Agents.

Then the usage CSV landed, and the last twenty-four hours looked less like a workday and more like a small storm that decided to live on my laptop.

The numbers (from the export, not vibes)

In the window ending ~9:11 PM IST on 18 Sep 2026:

• ~1.25 billion tokens burned
• ~95.7% cache reads — most of the heat was rereading context, not inventing it
• ~7.6 million output tokens actually written
• 214 distinct Cursor Cloud Agents touched the account
• Busiest five-minute slice (~8:30 PM IST): 84 agents still showing token activity at once
• Desk log of agents actually RUNNING at peak: about 70–85 concurrent (never ResourceExhausted on that climb)
• One session counter flashed ~141 launches — that was cumulative probes, not agents running together
• Almost all of it rode grok-4.6 xhigh-fast, with the desk bots chewing the rest
• Cost column on the export: Included / Free — no on-demand dollars in the file

For scale, and only as a footnote: that single day was about a fifth of all tokens in the longer export (~6.45B across ~29 days). The story stays yesterday evening.

I almost wrote "100s of parallel" because the evening felt that loud. The receipts disagree in a useful way: seventy to eighty-five agents running at once, eighty-four still talking in the hottest five minutes of the CSV, two hundred fourteen named across the day. The honest flex is the swarm that stayed healthy — not a round number that did not survive the logs.

What the storm was doing

This was not one big feature. It was a factory finding its rhythm.

OrderEasy came back off park. We ramped agents one by one until capacity stopped complaining, then pushed again with QA executes sitting beside builders, then sped up Jira chaining so the slots never starved. By evening the open PR lists looked absurd on purpose:

• Backend (RetailerCustomerPlatform): on the order of ~40 new PRs in the burst, ~99 open on the tip
• Retailer web: ~80+ new thin PRs in the same window, 100+ open
• Customer web: ~60+ new thin PRs, ~70 open
• Scanner (buyeasy_retailer_scanner): live again — wiki/help PR already moving
• Delivery partner: public repo live — https://github.com/vin8003/ordereasy_delivery_partner — greenfield scaffolding already moving
• Jira OE project: 59 issues created in the last day

The tickets themselves are almost boring, and that is the point. Optional fields. Isolated serializers. Tiny FE probes that will not explode the stack when they land. Scout agents minting the next thin backlog while QA agents run cases on agent VMs with dummy data — never against live .win.

Somewhere in the same twenty-four hours, the other side of the desk kept the X sprint alive: morning originals for My Orders and Message shop, evening originals for operating hours, verified cites, and the Building now crop from vin8003.com. Replies stayed drafts on purpose. SuperGrok Heavy is a renewal decision, not a vibe.

What it felt like

There is a particular quiet that arrives when you tell the bots to stop narrating.

Earlier I asked for less chatter — no merge-ready drip, no "still climbing" updates, escalate only if something is on fire or only I can decide. The agents kept working. The chat got quieter. The PR numbers did not.

That is the whole experiment in one sentence: orchestrate hard, speak soft, let the factory show its work in Git.

I am not going to pretend every PR is a masterpiece. Some are probes. Some will die. Desk still does not merge OrderEasy for me — Done is still a button I have to press. The honest scoreboard is not "shipped to production tonight." It is "how much parallel, isolated work can one founder desk keep honest before the tools buckle."

Tonight the tools did not buckle at ten. They did not buckle at twenty. The usage file and the running-agent log agree close enough: about seventy to eighty-five agents working at once, more than two hundred named across the day, and the tools did not fall over.

That is enough of a glimpse.

Closing

If you are reading this as a lab note: the artifact is the CSV, the PR lists, and the quiet lock.

If you are reading this as a founder: the interesting part is not the billion tokens. It is that most of them were cache — the system remembering what it already knew — while a swarm of narrow agents each owned one small, readable change.

One job each. Same thesis as the site. Just louder, for one night.