On 24 September 2026, Y Combinator CEO Garry Tan posted exactly two lines: Make software agents want and Use agents to make people want software. The first line is about selling to machines. The second is about letting machines sell to people. If you run a small business, the question is not which line is right. The question is which one you start with this week.
The short answer: Vector A means making your software something an AI agent can find, connect to, and buy from. Vector B means putting agents to work on marketing, content, and lead follow-up while humans still approve and pay. A Vietnamese SMB should run Vector B now, because it needs no API and it overlaps with the SEO and GEO work already on your desk. Add Vector A only once your product has an API.
None of this appeared out of nowhere this week. In August, Garry Tan quote-posted Firecrawl's AnyDoc with the line Make something agents want. In March, Forbes had already framed B2A and the idea of Agent-Market Fit as an update to Y Combinator's famous mantra, make something people want. What is new on 24 September is that he stitched the two sides of distribution into two named vectors — and the early replies on the thread went straight for the tension: if agents market to agents while people just want the job done, the old funnel survives until agents actually pay.
This article reads both vectors through the eyes of a small-business owner, counterarguments included, because the thesis reads as pure hype the moment you leave them out.
Vector A: make software agents want — how Agent-Market Fit works
The canonical Vector A story: a coding agent is building an app and needs a host for image models. It finds fal.ai's MCP on GitHub, asks the developer one question — "want me to set it up?" — and two minutes later fal.ai has a paying customer. No landing page, no ads, no sales call. The buyer never opened the website.
Agent-Market Fit, as the 24/09 research thread lays it out, comes down to five conditions:
| Condition | What it looks like in 2026 |
|---|---|
| Discover: the agent can find you | MCP registries, complete documentation, training data, llms.txt |
| Parse: the agent can read you | OpenAPI, clean markdown, error messages a machine can act on |
| Connect: the agent can plug in | API key in under two minutes, OAuth, a sandbox separated from production |
| Pay: the agent can pay | Usage pricing a machine can read; a "contact sales" button means a lost customer |
| Trust: the agent can rely on you | Idempotent tools, evaluable output, no unintended side effects |
Reality is harsher than the checklist. A March 2026 survey of 1,456 SaaS products put average agent compatibility at just 36 out of 100. OpenAPI: 68.5%. MCP: 2.2%. llms.txt: 0.5%. Products that shipped an MCP tool scored 17.5 points higher. Stripe and Supabase led the field at 58 each. That spread is the market gap: almost nothing is agent-ready, while the MCP leaderboard already lists 132,632 servers and grows by roughly 19,000 a month. The distribution surface is turning into an app store for machines — and in any app store, being listed has never meant being chosen.
Two voices worth hearing. Harj Taggar's read is that devtools are growing because the power user is now an agent, and agents write more code than people do. The founder question he leaves hanging is a good one: what will agents decide on a human's behalf, and what infrastructure does that decision need to be right? Savneet Singh tells it from the operator's side — restaurant operators don't open dashboards. They tell the agent: "find out why food cost jumped, and fix it if you can." The agent picks the system with a clean API and actions it can take, not the one with the prettiest interface.
Gartner (July 2026) sizes the shift at $234B of application spend between 2026 and 2030 exposing agentic arbitrage. Agents work across many systems at once, which strips seat licenses and dashboards of their reason to exist — roughly 20% of SaaS spend by 2030.
But Vector A has hard edges. An agent is not a cardholder, does not sign contracts, and does not carry SOC 2 responsibility. The thesis holds for API-first products, devtools, infrastructure, and payments. It breaks against traditional enterprise procurement, tightly regulated industries, and products whose only surface is a UI for the owner.
Vector B: use agents to make people want software — where the money already is
Vector B is the one that concerns you whether you sell kitchens, logistics, or software. Don't collapse it into "marketing agents." It has three separate layers:
| Layer | The work | Hard signal |
|---|---|---|
| Discovery: get cited by AI | GEO/AEO — being named by ChatGPT, Perplexity, and AI Mode | G2 2026: chatbots are the number-one source for B2B shortlists; 69% of buyers switch vendors based on AI; one in three buys from a company they had never heard of; under 10% of AI-cited sources rank in Google's top 10 |
| Manufacturing demand | Content, directories, newsletters, and short video, drafted by agents | Cody Schneider: a directory hitting 100 clicks a day after 90 days, a list of 20,000 in six months. Greg Isenberg: 23 loops wired into Stripe, PostHog, and G2 |
| No lead left behind | SDR work, B-lead nurture, inbound handled by agents | SaaStr: 3 people plus 20 agents, year over year from −19% to +47%; a sales agent working 1,000 forgotten warm leads closed $2.7M; Amelia reached 614 meetings and 402,000 chats and events for about $257 a month in compute |
The G2 cluster is the part that should worry anyone who works in search: chatbots have overtaken Google as the first place B2B buyers build a shortlist, and fewer than 10% of the sources AI cites rank in Google's top ten. Doing traditional SEO well does not guarantee you get named. That is why GEO (Generative Engine Optimization) exists — and why this blog's Local SEO cluster exists. If you are starting from zero, read Local SEO with AI Overviews for Vietnamese businesses first.
The market is not waiting. In the week of 14–22 September, OpenAI opened Sponsored Agents in alpha in the US with connectors for HubSpot and Shopify. The same week, HubSpot shipped Agent Hub with four agents covering Campaign, Content, Nurture, and Revenue. When agent-run GTM becomes a feature of the CRM you already pay for, it stops being a startup trick and becomes the new default way of operating.
The cleanest part of Vector B is the word incremental. Jason Lemkin put his finger on the right place to point it: don't let AI poach your A-leads, the ones that deserve a human. Let agents eat the B-leads your team never has time to follow up on. That is genuinely new revenue, not revenue moved out of a human's pocket into a machine's.
Vector B's bottleneck is not the model — the model is good enough. The bottleneck is taste, brand voice, the approval step before anything goes out, and measurement: AI-referred traffic has to be tracked on its own, never folded into organic. That is exactly the weekly content, SEO, and GEO work at 5ac, and you can see how we log it in AI agents and local SEO for Vietnamese SMBs 2026.
The two vectors are not opposites — they are one flywheel
Reading Garry Tan's two lines as either/or is reading them wrong. The vectors feed each other:
- A product that solves a real job is the floor under both.
- Vector B brings people in: they learn, they want, they compare.
- Vector A brings agents in: other agents find your tool and connect on their own.
- Humans pay and approve; agents generate usage data, cases, and evaluations.
- That data makes the product better, and the loop starts again.
Without a real product underneath, both vectors are theater. MCP will not rescue a tool that fails the moment an agent calls it. A content farm will not rescue something a customer can rebuild in an afternoon with the same model. Make people want is still the binding constraint; the two vectors are just two faster roads to it.
Three counterarguments to read before you act
A balanced read has to include these, because the thesis has three genuine weak points.
One: agents marketing to agents. The early replies on the thread point at the paradox — people do not want to "manage a team of agents." They want the work finished. The old funnel is not dead; it survives until agents actually sign and pay, which has not happened at scale.
Two: humans still sign. Agents do not pay, do not carry legal liability, and do not pass security audits. Every revenue figure in the Vector B section has a person at the end of the funnel approving and paying. Remove the human step and you remove the revenue.
Three: sprawl. Gartner (July 2026) forecasts that by 2028 agents will outnumber salespeople 10 to 1, yet fewer than 40% of sellers report higher productivity. Creating agents is easier than creating results; a company that deploys 20 agents without measuring anything is multiplying chaos, not revenue. Wired, through Andy Walters' analysis, leans the same way.
What an SMB should actually do this week
Turn the two vectors into concrete work, in priority order:
| Priority | What to do | How to start |
|---|---|---|
| Now (Vector B) | GEO: become a source AI cites | Every week, ask 10 typical buyer questions through ChatGPT, Perplexity, and Gemini; log who got cited and why |
| Now (Vector B) | Content drafted by agents, approved by people | Target specific questions like "how much does an acrylic kitchen cabinet cost in Go Vap", not generic "we have AI" copy |
| Now (Vector B) | Nurture B-leads with agents | Take the old lead list nobody touched, have agents draft segmented messages, and have a human approve before anything sends |
| Measurement | Separate the AI-referred channel | Track AI-search traffic on its own instead of blending it into organic; let that channel decide your content plan |
| When you have an API (Vector A) | Make the product agent-ready | OpenAPI and a real error model, remote MCP with a separate sandbox, a /.well-known/mcp.json file; then test it by telling Claude Code or Cursor to "find a tool that does X" and see whether your name comes up |
| Avoid three things | Skip the theater | No empty MCP published just to appear in a registry; no autonomous cold email for a furniture brand; no "multi-agent platform" landing page when customers are buying a finished job |
If you want an operating frame for agents drafting while humans hold the approval step, AI CRM sales workflows: from prompts to pipeline walks through it step by step. And for the thinking underneath — owning your own agents instead of renting them — see the earlier piece on Garry Tan and personal AGI.
Conclusion: two lines, two roads, one final constraint
You can read Garry Tan's two lines in either order, but you cannot do the work in either order. For a small business, Vector B is this week: content and GEO drafted by agents, approved by a human, with the AI-referred channel measured separately. Vector A is the next phase, once your product already has an API agents can connect to. Both vectors feed one flywheel, and both answer to the same final constraint — at the end of the day, someone still has to want it and someone still has to pay.
Your competitors read the same two lines. What separates you is who starts on Monday.
Frequently asked questions
What are Garry Tan's two GTM vectors?
On 24 September 2026, Y Combinator CEO Garry Tan posted two lines: Make software agents want, and Use agents to make people want software. The first vector is selling to agents through APIs, MCP, and llms.txt. The second is putting agents to work on marketing, content, and lead follow-up, with humans still approving and paying.
Which vector should an SMB start with?
Start with Vector B — using agents to run marketing and nurture leads — because it needs no API, it can begin this week, and it overlaps with the content and SEO/GEO work you already do. Add Vector A only once your product already has an API that agents can connect to.
What does Vector A require to make a product something agents want?
Five conditions: agents can discover you through MCP registries, real documentation, and llms.txt; they can parse you through OpenAPI and clean markdown; they can connect with an API key in under two minutes, with a sandbox separated from production; pricing is usage-based and machine-readable; and tool actions are idempotent, evaluable, and free of unintended side effects.
How do you measure whether Vector B is working?
Track AI-referred traffic separately instead of blending it into organic. Every week, ask 10 typical buyer questions through ChatGPT, Perplexity, and Gemini and log who gets cited. Then watch the B-leads nobody used to follow up on, because that is the incremental revenue agents unlock.
Will agents fully replace human marketing?
Not yet. People still approve content before it ships, still sign contracts, and still pay. Gartner warns that by 2028 agents will outnumber salespeople 10 to 1 while fewer than 40% of sellers report higher productivity. The real bottleneck is not the model — it is taste, brand voice, and the approval step.
Sources: the figures and quotes in this article are drawn from 5ac's AI/Tech Radar research thread of 24/09/2026, whose own sources are Garry Tan's thread on X (24/09/2026), Forbes on B2A (03/2026), a Gartner report (07/2026), a G2 buyer survey (2026), SaaStr, the MCP Toplist (23/09/2026), and a March 2026 agent-compatibility survey of 1,456 SaaS products.
Want a team of agents drafting your content while a human keeps the approval step? 5ac.vn builds and runs agent systems for small businesses every day. See G-Company OS pricing to start with the plan that fits.
Related articles
- Local SEO with AI Overviews for Vietnamese businesses
- AI agents and local SEO for Vietnamese SMBs 2026
- How Vietnamese businesses get cited in AI search
- Garry Tan: own your intelligence with personal AGI
- AI marketing for small business: workflows that grow leads
- AI CRM sales workflows: from prompts to pipeline
- AI agent trends 2026 for Vietnamese businesses