#203 — Agent-led growth (ALG) for startups

Agent-led growth is a go-to-market motion where AI agents, not humans, drive discovery, evaluation, and purchase decisions on the buyer's behalf. PLG made the product itself the sales channel for humans; ALG makes documentation and APIs the sales channel for machines. Supabase went from 1 million to 4.5 million developers in under 12 months largely through agent selection in tools like Bolt and Cursor, not a PLG funnel.

WebMCP (Web Model Context Protocol) is an open standard Google and Microsoft shipped in Chrome 146 on February 12, 2026, that lets AI agents interact with your existing web forms directly instead of screenshotting and guessing where to click. You implement it by adding a tool name and description as HTML attributes to forms you already have — no backend rebuild required. If your buyers use AI agents to evaluate software (94% of B2B buyers already do), skipping this means agents can't test your product properly.

Token-to-value counts how many tokens an AI agent burns before it confidently recommends your product, then how many more to implement it. You estimate it by running your own docs through Claude Code or a similar agent and timing how long it takes to reach a working integration — Resend gets there in one command, which is why Claude Code picks it over SendGrid 63% of the time versus 7%. Documentation gaps, vague pricing, and missing code examples all inflate this number.

Claude Code selects Resend 63% of the time versus 7% for SendGrid because Resend has cleaner, more machine-readable documentation, a free tier that skips budget approval, and a single-command implementation path. SendGrid is the larger, more established company, which shows agent selection doesn't follow brand size or incumbency the way human buying does.

There's no fixed page count, but the benchmark is whether an agent can answer every question a sales call would surface without contacting your team. Supabase and Resend both invested more heavily in developer docs than in traditional marketing, and both became default picks in AI coding tools as a result. Audit your docs specifically against token-to-value, not just against whether version numbers are current.

Usage-based or free-tier pricing removes the budget approval step that blocks an AI agent from provisioning your product on a buyer's behalf. Supabase and Resend both use this model, which is part of why agents can select and implement them without a human sales cycle. If your pricing requires a quote or sales call before a trial starts, agents can't complete the evaluation-to-purchase loop for you.

No, Generative Engine Optimization (GEO) only covers the discovery stage — getting cited when an agent is problem-aware. ALG extends through evaluation (agents building feature matrices and testing capabilities), decision, and purchase, where 77% of buyers already purchase from whatever their AI flagged as favorite. GEO is table stakes for ALG, not the whole strategy.

No, but the seller's role shifts from creating demand to confirming or contesting a decision the agent already made before the sales call started. This raises the stakes of remaining human interactions rather than eliminating them, since agents reward consistency between marketing claims, documentation, and what sellers say on calls. ALG works as a layer on top of existing GTM, not a replacement for it.

Developer tools tipped to agent-led defaults in under 18 months; categories like CRM, HR software, marketing automation, and vertical B2B SaaS haven't crowned their defaults yet. The open question is which of these categories reorganizes first and whether current leaders adapt fast enough to earn default status before a Supabase- or Resend-style challenger takes it from them.

Focus on three properties: findable (invest in GEO/AEO so agents surface you), evaluable (publish documentation that answers every question a demo call would), and actionable (work with engineering to annotate your pricing, trial, and demo pages for WebMCP). You don't need a fully agentic product; removing friction at each of these three stages is enough to start competing for agent recommendations.

Anthropic's Model Context Protocol (MCP) lets AI agents connect to backend tools and data, while Google's Agent-to-Agent protocol (A2A) lets agents coordinate with each other. Together with Google/Stripe's agent-payment protocols and the newer WebMCP browser standard, they form the four-layer stack that agent-led growth runs on.

Instant Checkout is OpenAI's feature letting ChatGPT complete a purchase directly inside the chat, currently focused on consumer transactions. It's an early signal for B2B software: agent-initiated purchase at the point of recommendation is near-term, not hypothetical, which is why pricing pages need to be agent-actionable now.

SEO optimizes for human search behavior and ranking algorithms; GEO (Generative Engine Optimization) optimizes for getting cited when an AI agent is researching on a buyer's behalf. GEO is described as table stakes, not a full strategy — the balance between SEO and GEO investment is a live question teams should revisit regularly rather than set once.

Google and Microsoft published the WebMCP standard on February 12, 2026, and it shipped in Chrome 146. It closed the gap where agents previously wasted two-thirds of AI-web interaction compute screenshotting pages and guessing where to click.

Supabase went from a $765M valuation in September 2024 to $5B by October 2025, driven largely by becoming the default backend inside vibe-coding tools like Bolt, Lovable, and Cursor. Its developer base grew from 1 million to 4.5 million in under 12 months over the same period, without a traditional BD-led sales motion.

The developer ecosystem tipped to agent-selected infrastructure defaults in under 18 months. Category-specific B2B SaaS — CRM, HR software, marketing automation, vertical software — hasn't crowned its defaults yet, which is the open window founders in those categories still have.

These tools evaluate infrastructure options based on documentation quality, API predictability, and ease of implementation, then make the selection on the developer's behalf rather than presenting a menu. This is how Supabase and Resend became defaults — through repeated agent selection across thousands of sessions, not through a sales pipeline.

94% of B2B buyers are already using LLMs in their research process, and 77% end up purchasing from whatever their AI-informed evaluation flagged as the preliminary favorite. This means a buyer's agent frequently forms a point of view before a seller ever gets on a call.

A machine trust moat forms when agents that find your tools reliable start preferentially recommending you over time, similar to how repeat positive experiences build human brand loyalty. Unlike brand trust, it compounds through technical signals — documentation quality, API reliability, low token-to-value — rather than marketing spend or awareness campaigns.

Yes, you don't need a fully agentic product to benefit, the same way companies adopted parts of the PLG playbook without becoming Dropbox. Removing friction at the findable, evaluable, and actionable stages captures most of the upside even with partial implementation.