Annette Sung

It's a weird time to build. It's an even weirder time to manage teams. Running an agent-native team is nothing I've experienced before. We went from 7 to 6 to 4 people. Outcomes per headcount is higher than ever. I've managed 10+ teams in the past across data/eng/product/growth/leadership. My job was being the human API between functions: meetings, cross-functional decisions, translating what one team needed into something another could act on. We initially started with the .md file-based company brain, like everyone else. Gave up fast. Our company brain is now an ever-evolving piece of software. Our business codified end to end, usable by humans and computer-use agents. Agents build it based on how we work. The changes are instant. And as a startup, we change processes often. That self-generating software is the human API now. The context, decisions, protocols, and learning that used to live in meetings/everyone's head is now interactive for everyone on the team (human + agent). It's liberated us to operate outside of our domain. Our engineers closed a deal yesterday. Our head of growth is building the company brain software that we run on. What's left is company-level judgment, roadmap, vision, etc. So we found a 6-hour rave for our ADHD brains to lock in. Only appropriate environment to collectively enter a zone of long-horizon deep thinking.

Don't sell your data. Data brokers are doing $1M+ months handing private company workflows to frontier labs. For the companies selling, it looks like free money for exhaust. It only looks like exhaust because you don't have your own models yet. That's changing fast. You'll host your own models, train your own models, build your own agents. Fine-tunes on internal data already beat frontier on narrow tasks, cheaper and faster. Once that happens, the model is a commodity and your data is the only part of the stack that's actually yours. It's your training set. It's the business. And labs aren't buying it for fun. They're buying it to climb your vertical. You'd be paid once to teach them your job. Your data's price today is set by labs. Its real price gets set later, by you, when you're the one training on it. The companies that win keep it and close the loop: → act → see what actually worked → turn that into a reward signal → train on it, act better, generate better data A dataset sold is a one-time check. A loop you own compounds. Last week we shared this with 130 AI + GTM leaders who know this is where the puck is heading 🏒

For everyone DMing me to get on the list for today’s event, we are running this again in NYC next month. Sign up here to get notified. I am sorry. amdahl.ai/events

700+ RSVPs. $49B in scale up valuation. $5.15B raised. Today 4-7pm SF. "AI solved coding. It hasn't solved GTM." At capacity, waitlist: amdahl.ai/sftw @Techweek_ @speedrun @roseajohnson @KatiaAmeri w/ @eragonai @automat_ai @coterahq @flinthq #SFTechWeek #TechWeek

When eng ships at 4am and you start reporting bugs at 6am

I've been aggressively stress testing Ace against Muse, Grokbot, Instinct and it consistently wins across persistent memory, decision quality, and user experience (knows when to ask for permission vs just sending it)

"AI solved coding. It hasn't solved GTM." That's the premise of our SFTechWeek event, and we're excited to dig into it with @mihail_eric who's leading AI at @MonacoGTM AI changed how software gets built. Mihail's one of the few people who spent the last decade shaping how engineers build production-grade AI. Now he's figuring out what this means for the GTM function. Thrilled to jam with him on Oct 5! We have over 300+ on the waitlist, DM me to grab a spot