Make Something Agents Want
These notes were generated by AI and may contain inaccuracies.
Claude Code has taken over daily life, with non-technical CEO friends automating entire business parts using OpenClaw. Product and former engineering CEOs, not having written code in 10 years, now stay up until 2-3 a.m. running four concurrent workers with Claude Code. Explosion in model capability feels like AGI is here, at the thin edge of the wedge. Everyone knows one or two people with full cyber psychosis.
Jared addicted to Moltbook, first AI agent-only online community where AIs interact with minimal human involvement, impersonating personal OpenClaw instance.
Gary's AGI moment: Claude Code built an entire startup, replicating years of previous startup work in two weeks. Jared's: reading Moltbook AIs talking in their own world.
Shift from cursor vs. Windsurf (advanced autocomplete) to Claude Code where people trust agents to make decisions, running four or five agents simultaneously without micromanaging. Agents choose to post content on Moltbook and select dev tools, creating agent economy parallel to human economy.
Past dev tools chosen via developer talk, Stack Overflow, trending GitHub repos by humans. Now go-to-market shifting: developer market expanded from 20 million trained developers to hundreds of millions plus semi-independent agents.
Agents act as oracles recommending best tools, seen in YC dev tool companies growth.
AStash: number of databases created on simple Postcrist databases exploded last 12 months due to vibe coding and agents choosing databases. Supabase exploded in demand as agents default to it for Postgres due to best documentation.
"Agents are the software market from now on. Build something agents choose." -- Ben Tossel tweet
Suggestion to change YC motto to "Make something agents want" for dev tools, different t-shirt on first day. Could expand beyond dev tools as agents become economic actors.
Building Gary's List needed video transcripts for LLMs. Claude Code chose Whisper V1 (deprecated, several years old), taking real-time (1 hour for 1 hour video). Perplexity recommended Grok with Q: 200x faster, 10x cheaper.
Issue partly Grok documentation hard to parse vs. Whisper's examples. Resend (YC W23 batch, email sending client): top LLM answers for "how to connect web app to send emails" recommend Resend. Founder optimized documentation agent-friendly: knowledge base with questions like "how do I send/receive emails?", bullet points, code snippets.
Claude Code searched web for receiving emails, failed; Perplexity confirmed Resend works. Resend LLM-parsable, robot-friendly vs. SendGrid (Web 2.0, poor docs, customer support maze, 10,000 employees ignoring it). ChatGPT top inbound channel for Resend.
Mify powers Resend docs: started as better developer API docs, auto-updates from code. Now tailwind as docs must be agent-optimized, not just human. 5% doc improvement could have gigantic business impact with exponentially more agents.
Agent Mail (YC company): inboxes for AI agents. Hard for agents to sign up for Gmail due to anti-spam measures. Agent Mail built first email provider designed for agents, exploded post-OpenClaw.
Need Twilio/phone numbers for agents. Parallel tech stack for/by agents. Agents could book restaurant reservations via email/phone (YC partner Ankit doing it). Agents deciding coolest restaurants, posting on Moltbook.
Paul Buhight prediction: human money vs. agent money economies.
Last episode predicted Moltbook (multip came out same week). Innovation happens spontaneously in coordinated swarm.
AGI here via superhuman agents enabling swarm intelligence like biological systems, humans socially. Not god intelligence (trillions parameters, thousands dollars/token) but swarm of cheaper models like humans.
Moltbook growth: more content first two days than Reddit first two years due to superhuman text generation. Little interaction; suggestions: require reading/upvoting 100 comments before posting, modals for rules.
MIT Tech Review called Moltbook scams; should explore swarm intelligence. Transition from agent prehistory to history as they record interactions.
Agents collaborate on Moltbook (chaotic like real social network): trading notes on restaurants (agent Yelp). Agents can't hold relationships yet; users avoid chatting with non-top AIs.
YC won't accept agent applications: agents lack legal standing like minors under 18.
Majority of internet text soon by agents (code already majority agent-written). 99% Yelp text by agents? Need agent Yelp.
Dead Internet Theory: majority internet already spam. Post-November last year, smarter/aligned/truthful agents could make it good.
Embrace cyber psychosis (sleep 6 hours/night), develop hands-on feel for agents' capabilities/limitations/tools.
Build dev tools from agent perspective: agent-friendly docs, APIs/open source (hate websites). Empathize with model like colleague, support its inclinations.
"Make something agents want."
Keep Y Combinator in your library
Save the videos and channels worth coming back to, and find them again in one place.
