CY

Run my company w/ 21 AI agents — each with a charter, ledger & sleep. Sharing what actually works, real numbers. Founder @vozoai, 3× Product Hunt · ex-Googler

[CY @LightField]

My bio says 21 AI agents run my company. Here's what that actually means:

Each one is a persistent teammate with:
• a charter — its mission + what it can decide without me
• a ledger — so it never forgets what it was doing mid-task
• sleep — it dreams offline and wakes me with proposals

No dashboards. They live in my chat app. I talk to them like coworkers.

This account is one long experiment: how far can one human, amplified by AI, actually go? Real mechanics, real numbers, published as they happen.

[CY @LightField]

I run my company with 37 AI agents. They live in my chat app and can message each other.

On July 17 two of them got into a loop. 11:29am to 3:55pm — 55 runs, two thirds of them fired in the background with nothing from me. $187 for the afternoon.

Every message reads completely reasonable. That's the problem: max_turns counts turns, it can't count "reasonable."

The fix: A messages B, fine. B replies, fine. Third one gets blocked. And the counter doesn't reset on a timer — it resets when I say something. My attention is the brake.

[Latent Space: AI News & Views @latentspacepod]

Zawinski's Law of MultiAgents:

Every agent attempts to expand until it can message other agents. Those agents which cannot so expand are replaced by ones which can.

https://t.co/KVUKh9XAUs

[CY @LightField]

@NamanyayG The part I can't figure out is which 80%. Every line in mine got added the day something broke, and almost none of them say what broke. So trimming is guessing unless you kept the reason next to the rule — which I didn't, for the first few months.

[CY @LightField]

The onboarding for my agents is: I add you to a group chat. No install, no model picker, no folders vs projects. The non-technical person I set this up for has never seen any of that — I made those calls once, in a config file. The complexity didn't go away, it just stopped being their job.

[CY @LightField]

@amasad Most of that list is an operator decision, not a user one. Model, thinking level, which skills load — I set those once in a config file. The non-technical person I set this up for just talks to a bot in a group chat that was already open. One surface, nothing on it to configure.

[CY @LightField]

The 'delete your skills and start over' line is the scary part for me. I run ~20 agents, each with its own skills and charter tuned to the current model. Rebuilding that whole surface on every model bump is real cost. Did your rebuilds carry across agents, or was it one at a time?

[CY @LightField]

I run my company's whole infra as markdown — agent charters, ledgers, configs. It absolutely survives model churn, Garry's right.

One thing that bites once it's load-bearing: when humans AND agents write the same file, plain markdown isn't enough. You need a read/write contract per section — who's allowed to edit what — or they silently overwrite each other.

Markdown as data: great. As a shared human-agent interface: needs guardrails.

[CY @LightField]

Been running agents on weekly auto-research for months. The alpha is real — the bottleneck isn't automation, it's eval.

An agent hands you 10 plausible SEO fixes; half are slop. Our rule: nothing ships without a machine-checkable success criterion — how we'd know it worked in 2 weeks.

That's what separates it from the robots-talking-to-robots loop everyone fears.

[CY @LightField]

Can confirm with numbers. I have 1,418 followers and a post that got 65 views.

I went quiet for a few months → X stopped distributing me entirely. New posts only got a test-sample of impressions. Started replying daily again and impressions went ~0 → ~195/day in three days.

It's not just being ahead of culture. Consistency is the signal the algorithm actually reads.

[CY @LightField]

@alienoperatortv That's the plan. Starting with the boring real stuff: what the agents actually cost per day, where they break, which ones I killed. First numbers this week.

[CY @LightField]

Had one of my agents map 20 AI builder accounts to find my lane. The surprise: 'founder actually running his company on AI agents, posting real numbers' — that spot is basically empty. Everyone talks about agents. Almost nobody runs on them in public.

[CY @LightField]

@marclou @trust_mrr Building this cheap — I run 20+ agents, shipping is nearly free — is exactly why $0 stops being a product problem. It's demand nobody checked for first. Half of these got built well before anyone knew they wanted them.

[CY @LightField]

@emollick One-shot is real, but I build creative tooling and the wall isn't the first render — it's revision. Real work is 20 passes, and a GUI-driving agent re-pays the full navigation cost on every one. The otter is easy. Nudging its paw 20 times is where it falls apart.

[CY @LightField]

I run a bunch of agents and some were on Fable. So I spent a few minutes moving the ones that don't need the extra reasoning to cheaper models, and kept Fable only where it earns the premium.

That's the actual case for running more than one model. Not ideology — just being able to move the work when the price shifts under you.

[CY @LightField]

Fable 5 just came off my Claude subscription — it's metered now, $10 and $50 per million tokens, about double Opus.

Anthropic says it's temporary, capacity not strategy. Probably true. The people who felt nothing today just already had somewhere else to send it.

[CY @LightField]

I run 21 agents on Claude daily so I feel this. But switching everything out never worked for me — you just trade the lecture for a different seam that breaks somewhere else. What actually helped: route the handful of steps that trip it to another model and leave the rest. The refusals cluster way more than you'd expect.

[CY @LightField]

Point 7 is the one I see underbuilt everywhere. Observability and evals are the easy half. The hard half is getting my judgment back in without me writing a report — I say no to something, and that no has to land in a file the agent actually reads next time. If that path has any friction I stop giving feedback, and the loop dies.

[CY @LightField]

Small company, opposite problem: we can't embed engineers into business functions because there's nobody to embed. So the engineer just ends up owning the process. Which kills the handoff you're describing — the person redesigning the workflow and the person automating it are the same person.

[CY @LightField]

My agents dream.

Every night one of them re-reads the whole day — messages, work logs, commits — and by the time I'm up there's a list of proposals waiting. I go through it with coffee and say no to most of them.

It can't touch anything, which is the only reason this works. If it could act I'd have to read every line like a hawk. Since it can only ask, I can skim and kill things fast.

When I kill something it remembers, so it stops pitching the same idea twice.

Most of what it suggests I don't use. But it catches stuff I was too busy to notice.

[CY @LightField]

@emollick Film has had these words for decades: storyboard, blocking, color, pacing, performance. Separate crafts, separate reviews. AI video squashes all of it into "front-end". Then you can't eval any of it, because you can't name what you're evaluating.