Finding signal on Twitter is more difficult than it used to be. We curate the best tweets on topics like AI, startups, and product development every weekday so you can focus on what matters.

Your product’s next million users won’t have eyes

Aakash Gupta

Your product’s next million users won’t have eyes. Andrej Karpathy said it last week: “It’s 2026. Build. For. Agents.” He was quote-tweeting a Polymarket CLI that lets autonomous software access prediction markets through the terminal. Most product teams have zero surface area for this. MCP went from zero to 97 million monthly SDK downloads in its first year. 10,000+ active servers. OpenAI, Google, Microsoft, Cloudflare all adopted it. Anthropic donated it to the Linux Foundation because the standard had already won. Gartner projects 40% of enterprise apps will embed AI agents by end of 2026, up from less than 5% in 2025. Every generation of software has a dominant distribution channel. Retail (Microsoft won shelf space). Web (Salesforce killed on-premise with a URL). Mobile (Instagram ate companies that tried to port desktop). AI Discovery (if ChatGPT describes your competitor, you lose the click). Now: Agent Distribution. Agents discover through CLIs, MCP servers, and machine-readable docs. They don’t onboard. They connect, authenticate, execute, move on. The companies that build for the new interface first capture outsized share. The ones that retrofit lose ground they never recover. Three layers matter. Documentation (AGENTS.md, already adopted by 60,000+ projects). CLI (the entire Unix philosophy was accidentally designed for agents decades before they existed). MCP Server (Stripe lets agents review payments and process refunds, Cloudflare exposes 2,500 endpoints through just two tools). I talked to Todd Olson at Pendo and Brian Helmig at Zapier for the podcast. Zapier exposed 30,000 actions across 8,000 apps as MCP tools. Brian was honest: “It’ll work one time and then it’ll go off the rails the next time.” That’s why tool descriptions matter more than having a server at all. Engineers build the MCP server. But the quality of tool descriptions? Which features to expose first? Starting read-only? That’s product judgment. Stripe wrote “review payments, troubleshoot declines, process refunds.” An agent knows exactly what to do. Compare that to “manages payment operations.” The first gets picked. The second gets skipped. Research shows agents start failing at 30+ tools when descriptions overlap. I wrote the most practical guide I could on owning this shift: PM’s role, strategy meeting playbook, the first PRD template for agent access, a Claude Code sprint from zero to working MCP server, five production teardowns, and seven mistakes that kill adoption. The agents are already looking for your product. Whether they find it is up to you. https://news.aakashg.com/p/master-ai-age…

The PM's Guide to Agent Distribution: MCP Servers, CLIs, and AGENTS.md

The PM's Guide to Agent Distribution: MCP Servers, CLIs, and AGENTS.md

Topics

We doomscroll, you upskill.

slop ⇢ substance ⇢ signal