We killed our flagship product on a Sunday. Monday revenue 3x'd.

For 14 months we'd been building what every YC partner dinner had told us to build: an "AI agent platform" that did everything. Workflows. Memory. Tool-use. A no-code canvas. A marketplace.

We had 320 logos on the homepage. We had a $40 ARR per seat that nobody complained about. We had a Slack with 4,000 developers who told us they loved us. We were also growing 6% month-over-month and had eight months of runway.

The quiet truth was that nobody used 80% of what we'd shipped. Our most-loved customer — a four-person ops team at a logistics company in Rotterdam — used exactly one workflow we'd built in week three. They paid us $2,400/year. Then they told their friends.

Meanwhile we were burning $112k/month on infrastructure for features twelve people had ever opened.

We started Halt because every "AI agent" demo I'd seen in 2023 broke the moment you handed it real customer data. The agents would loop. Hallucinate tool calls. Forget what you'd told them four messages ago.

Our first version was a 600-line Python script that ran one specific kind of workflow — extract structured data from messy inbound emails, route it somewhere — extremely well. We sold it for a flat $200/month and three customers paid in the first week.

Then we raised. The seed deck said "platform." Investors don't underwrite scripts; they underwrite categories. So we hired six engineers, built an SDK, built a UI, built integrations. Each layer made the original 600 lines harder to find.

By month ten the team was bigger than our active user count for any single feature outside email routing. We told ourselves the platform thesis would compound. It didn't. It just diluted.

The hardest part wasn't the money. It was the team meeting where I tried to explain that the dashboard half of us had spent four months on was probably going to be deleted.

One of my engineers — someone who'd left a senior role at Datadog to join us — asked, calmly, whether we even had a strategy or whether I just changed my mind every time I read a Substack post. I didn't have a clean answer. I went home and cried in my car for forty minutes in the parking garage.

The fundraising conversations were worse. Every existing investor wanted us to "lean into the platform." Two new investors had passed the week before specifically because we weren't enough of a platform. Our advisor told me, gently, that I was about to commit the most expensive form of cowardice: doing the thing the room expected.

I spent the next three weekends running a quiet audit. Every feature. Every customer. Every dollar.

On a Sunday in February I opened a fresh repo and started porting just the email-routing workflow back to the simple version. By midnight it ran. By 4am I'd deleted the rest of the product from a staging branch and rewritten the homepage to say one sentence: "Halt turns your messiest inbound emails into structured records. Nothing else."

Monday morning I sent the team a Loom explaining the change. Three people quit by Friday — fairly, I think. The remaining four shipped the new version to production by the following Tuesday.

Within a week, three customers who'd been on free trials upgraded to a new $1,200/month plan we'd posted on the pricing page. Within a month, a logistics company we'd never heard of paid us $36,000 upfront for a year. By the end of Q2, MRR had 3x'd from the prior quarter and gross margin had jumped from 41% to 78%.

The "platform" had been hiding the actual business inside it.

1. 1
   Default to subtraction.

   Every feature you ship is a tax you'll pay forever. We deleted 80% of the product and the remaining 20% paid for the company.

2. 2
   The investors who like your deck don't run your P&L.

   "Platform" is a fundraising word. It is rarely a customer word. Build for the second one.

3. 3
   Listen to who already pays you, not who might.

   Our four happiest customers were all using one feature. We were building five more for prospects we hadn't even met.

4. 4
   The cowardice tax is real.

   The thing you're scared to do — kill the product, fire the wrong hire, ship the smaller version — usually carries the highest expected return. The expected cost of waiting is invisible until it isn't.

5. 5
   Three days of focus beats three months of optionality.

   We rebuilt the company in a weekend. The previous fourteen months had been an elaborate way to avoid that weekend.