People ask what the day looks like, the day a founder decides to shut down. The truthful answer is: it does not look like a day. It looks like weeks of rationalisation followed by a small moment that finally tips the call. Across three shutdowns, the small moment was different each time. The underlying signal was the same.

This post compresses the three decision moments into one read. Not for catharsis. For pattern. The pattern is what I now use to set kill thresholds for Gravity in advance, on calendar, before the rationalisation has anywhere to live.

The shape of a shutdown day

The day starts normal. There is a build to ship, a customer to reply to, a metric to check. Some part of the founder already knows the equation has stopped working; that part has been writing a quiet rationalisation for weeks. Then a small thing happens that the rationalisation cannot absorb, and the equation becomes legible. The decision is made in the next twenty minutes; the announcement takes a few more days.

The day does not feel dramatic. It feels procedural. There is a ritual of writing the email to the team, drafting the note to users, calling the parents who fronted the early money, telling the wife or the partner before bed. The day after is the heavy one. The decision day is mostly relief.

MindWave, October 2023, the sales conversation

The day MindWave actually died was a sales call I was running myself. The prospect was a small mental-health practice considering MindWave for their patients. I was twenty minutes in, walking through the product, and I caught myself describing it as "gentler than the alternatives". The prospect nodded politely. I caught my own framing in real time and realised it was the third call that week where I had used the same line.

"Gentler" is a feature, not a 10x improvement. In a saturated category, gentler does not move switching behaviour. The full structural reasoning is in the mental health platform postmortem. The decision moment was the call. I closed the laptop, walked outside, and by the end of the walk I knew. The next two days were the announcement.

The trigger that day was hearing my own voice make the same insufficient pitch one too many times. The signal had been there for months in the conversion data; the trigger was auditory, not analytical.

Super AI, March 2025, the polite churn

The Super AI moment was a customer email. A long-time user, polite, technical, telling me she was switching to using GPT-4 directly because the routing layer was no longer adding enough value to justify the wrapper. She was right. The signal had been weakening for two quarters; the late-2024 model convergence had been visible in the cost-per-margin curve since November. Her email was the small thing that the rationalisation could not absorb.

The full structural postmortem is at Super AI postmortem; the deeper decision postmortem is at the mistakes I made with Super AI. The decision moment was reading her email twice and realising she had described the entire business model in a paragraph. By the next morning, the call was made.

The trigger was a single user articulating the structural problem more clearly than I had been articulating it to myself. That is a pattern: the user often names the problem before the founder is ready to.

Vibe AI, February 2026, the billing reconciliation

The Vibe AI moment was a monthly bill. I sat down on a Sunday to reconcile the month's foundation-model spend against the month's revenue, the way I had been doing for sixteen months. The loss was not unusually large; it was just the thirteenth month it had been there in the same shape. The threshold I had set quietly to myself, two consecutive months of cost-per-active-user above price, had been broken eleven months ago. I had been ignoring it.

The decision moment was opening the spreadsheet, scrolling to the cost column, and realising the rationalisation had been "scale will fix this" for a year. Scale had had a year. Scale was not fixing it. The full postmortem is at Vibe AI postmortem; the bootstrap-as-enforcement reasoning that came out of it is in why bootstrapping, not VC.

The trigger was a spreadsheet I had been looking at every month. The signal had been visible for a year. The discipline I lacked was a calendar-based check against an explicit threshold. That discipline is now the centre of how Gravity is built.

Signal-to-shutdown lag (months) Signal visible Months I waited after MindWave 3 mo lag Super AI 5 mo lag Vibe AI 11 mo lag Source: Aryan Agarwal, retrospective on three shutdowns, 2026.
Each shutdown lagged the signal. The lag is the rationalisation. Calendar-based threshold checks shorten it.

What actually triggers the call

Three triggers, one underlying mechanism. The trigger is whatever finally makes the rationalisation untenable; the underlying mechanism is always the same: an equation that has been losing for months. The trigger can be auditory (your own pitch), social (a user articulating the structural problem), or numerical (a spreadsheet you finally read carefully). It does not matter which.

The lesson for Gravity is not "watch for triggers". Triggers are downstream. The lesson is to write the threshold down in advance and check it on calendar, so the equation is never allowed to run for months while the rationalisation grows. The discipline is detailed in bootstrapping an AI agent platform in 2026.

How to not shut down late

Three rules I now hold:

The framework that emerged from three shutdowns is in three startups, three shutdowns. The three checks I missed are at the three checks I missed. The version of bet four that has these rules baked in is Gravity.

Frequently asked questions

When do you actually decide to shut down a startup?

The decision usually arrives weeks before the founder names it. The day you finally name it is the day a small thing tips the rationalisation that has been holding the project up. Across three shutdowns, the trigger was different each time. The pattern was the same: the equation had been losing for months and I had been describing it as a transition.

Was there a specific moment for each shutdown?

Yes. MindWave was a sales conversation where I caught myself selling on warmth instead of measurable outcome. Super AI was a customer who switched to the underlying model and told me politely. Vibe AI was a monthly billing reconciliation where the loss had compounded past the threshold I had quietly set. Three different triggers; one underlying signal.

What is the hardest part of the decision?

Telling the users who actually loved the product. The hardest emails I have written are not to investors or to teammates. They are to the people who were getting value from the product and who would now have to find an alternative. That is the part that lingers; that is also the part that gives the next product its conscience.

How do you avoid shutting down too late?

Set explicit kill thresholds in writing before the product ships, then check them on calendar instead of on intuition. The threshold I missed at Vibe AI was cost-per-active-user. The threshold I missed at Super AI was margin-per-capability. Founders who check thresholds when they feel concerned check them too late. The discipline is calendar-based.

Does shutting down get easier?

Slightly. The first shutdown felt like a referendum on me as a person. The second felt like a referendum on the strategy. The third felt like a structured review. The work of separating the founder from the product is hard the first time and slightly less hard each time after. Therapy and honest writing both help.

Three takeaways before you close this tab

Sources