The first 100 customers are a learning system
The first 100 customers are not a demand-generation target in the normal marketing sense. They are the smallest useful sample of the market: enough people to reveal repeatable pains, objections, words, channels, and buying triggers. That is why YC tells founders to launch, talk to users, and find customers who love the product. An agent can help with that work, but it should not replace founder taste. Use it to widen your search, keep the list clean, draft better outreach, and preserve every learning from the field.
Start with a narrow customer hypothesis, then ask an agent to turn it into observable signals. A signal can be a job post, a complaint in a forum, a repeated search phrase, a public tech stack, a newly funded company, a product review, or a founder asking for help. The agent should return prospects with the reason they match, the source URL, the pain it observed, and a recommended next action. The founder still decides who is worth contacting.
Use agents for the manual work, not the judgment
Paul Graham's point in Do Things that Don't Scale is still the right frame: founders make startups take off by manually recruiting users. Agents do not remove that obligation. They make the manual work cheaper to repeat. Instead of spending two hours searching Reddit, X, directories, and search results, you can ask an agent to prepare a daily prospect board. Instead of writing every email from a blank page, you can ask it to draft a specific message based on observed pain.
A useful first-100 system usually has five agents: a signal mapper, a community researcher, an outreach drafter, a founder-sales CRM, and a learning loop. The signal map decides where to look. The community agent finds high-intent conversations. The CRM agent keeps follow-ups honest. The measurement loop turns replies and calls into positioning changes.
What the agent should produce every day
Do not ask for growth ideas. Ask for an operating packet. A strong daily packet includes 20 to 40 sourced prospects, five conversations worth joining, three landing-page copy changes based on observed language, ten tailored outreach drafts, and a short decision log. The packet should separate evidence from recommendation. Evidence is what the agent found. Recommendation is what it thinks you should do next.
This distinction matters because agentic systems can plan, call tools, and collaborate across specialists, but the founder owns the risk. OpenAI's agent documentation frames agents around tools, state, approvals, and orchestration. For first-customer work, approvals are not bureaucracy. They are where you keep the brand from sounding automated, prevent low-quality outreach, and decide which prospects deserve personal attention.
The 100-customer operating rhythm
Run a weekly loop. Monday, refresh the customer signal map. Tuesday and Wednesday, run community and search discovery. Thursday, send the highest-quality founder-led outreach. Friday, review replies, calls, objections, and activation. Every week should improve one asset: the homepage, the demo script, the onboarding checklist, the offer, or the target segment.
The mistake is trying to automate scale before you have a message that earns replies. The better move is to automate preparation while keeping the founder in the conversation. Agents can help you reach the first 100 customers faster because they make the slow work visible, repeatable, and measurable.