The money moved. The jobs didn't follow.

Mangomint's VP Sales grew 7.2x in a market that Mindbody owned. By building less, not more.

Mangomint sells salon and spa software. They came into a market where incumbents had spent years adding integrations, admin panels, and feature checkboxes nobody asked for. The complexity wasn't accidental. It was the product of every reasonable short-term decision: add the requested feature, win the deal, keep the customer from churning to a competitor with a longer list. That logic compounds into software that requires training to use and support tickets to survive.

Mangomint consolidated. Scheduling, payments, client management, marketing on one surface instead of four tabs across three tools. Every feature you add creates a new decision point for the buyer: "do I use this?" Most of those decisions, after an enthusiastic onboarding week, resolve to no. That's complexity tax on your customer's attention. It shows up in churn, not in support tickets.

Here's where I keep getting stuck with this story. The lesson sounds clean in retrospect. In practice, building less is one of the hardest organizational bets you can make. Boards want roadmap momentum. Sales teams want feature charts. Investors want TAM stories that imply more surface area. Every legitimate pressure inside a growing company pushes toward addition. Staying deliberately smaller than you could be doesn't get easier when growth is working. It gets harder, because now you have more to protect.

The 7.2x didn't come from a product insight. It came from subtraction discipline: the organizational muscle to keep saying no when yes was easier, compounded over years. That's how you survive a concentrating market. You don't match every feature the well-capitalized competitor ships. You're the product that doesn't require a training program to use.

Claude Channels is going to get read as "pull is dead." Wrong read.

Pull isn't dead. What Channels signals is a direction: toward push-based, always-on, memory-rich agents that act without waiting to be asked. Most teams are still debating whether to adopt AI tools at all, while the architecture is moving toward "which agent told you first."

I've watched this play out in every AI tool rollout I've been part of. You deploy something useful. Train the team. Week one, enthusiastic. Week four, compliant. Week eight, everyone's back to the old workflow because it's frictionless and the AI tool requires someone to remember to open it. The activation energy problem kills more AI adoption than any technical failure. Channels inverts it. The tool shows up when the event fires (a calendar trigger, a pipeline flag, a deal going dark) instead of waiting to be summoned.

The deeper signal is Karpathy's self-tuning thesis. Right now, push means "we configured what to surface." Next: agents that calibrate from what you act on and ignore, getting sharper without explicit training. Channels is an early manifestation. The arrow, not the destination.

Here's where it gets hard. A push system that isn't calibrated becomes noise. Your team learns to ignore the channel, and you've built an expensive alert they've trained themselves to dismiss. The failure mode is signal quality: you've built the delivery mechanism before you've earned the interrupt.

Where this is heading: ambient agents that self-tune don't just solve the activation energy problem. Over time, they remove the human decision point. The loop tightens (agent surfaces signal, agent acts on signal, agent learns from outcome) until the human's role is ratification, not judgment. The question for operators is whether you're building the signal quality layer now, while you still control what the agent surfaces and acts on, or whether you're going to inherit whatever defaults ship with the platform.