Product-led growth made the product the salesperson. It assumed a person was on the other side of the signup page: someone who would start a trial, hit an aha moment, invite a colleague, and eventually swipe a card. Fifteen years of SaaS go-to-market sits on that assumption.
It is starting to fail. Your best prospect this quarter may never see your homepage. Their agent will. It will read your docs, call your API, compare your pricing to three rivals, and pick one. You will not know you were in the running. I call the response agent-led growth: a go-to-market model where AI agents sit on both sides of the deal. They are the channel through which customers find, evaluate, buy and use your product, and they are the workforce that acquires, activates and expands those customers.
The objection worth taking seriously
The strongest case against this idea is that it is AI SDRs with a new label, and the label is already taken. Lifesight's CEO Tobin Thomas introduced an agent-led growth flywheel in February 2025: autonomous agents running content, lead scoring, demos and customer success. Autoplay uses the term for an in-product agent that nudges users after signup. If agent-led growth only means automating the funnel, it is a tooling upgrade to the sales-led and product-led playbooks, not a new model.
That objection is right about the prior definitions and wrong about the opportunity. Both cover one face: agents working for the seller. Neither addresses the larger shift, which is agents working for the buyer. Netlify's Mathias Biilmann named the buyer side agent experience, but as a design discipline, not a growth model. Joining the two faces changes three things the narrower versions leave alone:
- Who you persuade. The first evaluator is often an agent that reads docs, tests APIs and compares pricing, not a person who watches a demo.
- What you measure. Signups and product-qualified leads undercount demand that arrives through an agent.
- How you price. When an agent does the work, seats stop tracking value. Outcomes and metered work do, the same move I describe at the top rung of the AI packaging ladder.
The second objection is that agents do not buy enterprise software. In regulated, relationship-heavy categories, that will hold for years: agents will shortlist and humans will sign. The model below is built for that gap. But in infrastructure, the buying has already moved.
Why now
Three shifts landed within about eighteen months, and together they break the PLG playbook.
- Agents became buyers. On Neon, the serverless Postgres company Databricks bought in 2025, the share of databases created by AI agents went from 0.1% in October 2023 to 80% in October 2025, according to Databricks' State of AI Agents report. Nobody at Neon sold those databases. An agent picked them.
- The rails exist. Stripe and OpenAI shipped Instant Checkout and the open Agentic Commerce Protocol on September 29, 2025, with a payment token that lets an agent pay without seeing credentials. MCP gave agents a standard way to call any product's tools. Gartner expects AI agents to intermediate more than $15 trillion of B2B spend by 2028.
- Go-to-market labor got cheap. Research, outreach, onboarding setup and first-line support are now tasks an agent can run at close to zero marginal cost. PLG removed humans from the funnel to cut acquisition cost. Agent-led growth puts high-touch service back in, delivered by agents.
Product-led growth made the product the salesperson. Agent-led growth makes the product legible to agents and makes agents the growth team.
The two faces of agent-led growth
Agent-Ready is the buyer side. The test: can an agent, acting for a buyer, get from “I need this job done” to “done” in your product without a human from your company?
| Capability | The test | What good looks like |
|---|---|---|
| Discoverable | Ask five leading agents to solve your core job. Do they name you? | Open docs, llms.txt, an MCP server listed in registries |
| Evaluable | Can an agent compare you without a sales call? | Machine-readable pricing and limits, sandbox keys, published evals |
| Transactable | Can an agent sign up and pay? | Programmatic signup, keys on creation, usage billing, agent checkout |
| Operable | Can an agent do everything a user can? | API and MCP parity with the UI, errors that say how to fix them |
| Trustworthy | Would a CISO let an agent in? | Scoped permissions, agent identity, audit logs, spend caps, approval gates |
Agent-Run is the seller side. Your own agents take a customer from first signal to expansion, with people at the moments that need judgment, trust or a signature.
| Stage | Agent job | Human role |
|---|---|---|
| Acquire | Research accounts, spot intent, send the first touch | Set the ICP, approve messaging |
| Convert | Answer technical questions, build a tailored sandbox | Negotiate and close large deals |
| Activate | Import data, configure integrations, finish the first task | Handle edge cases |
| Retain | Watch usage, fix issues before tickets exist | Own the executive relationship |
| Expand | Spot new jobs to be done, propose and provision the upgrade | Approve commercial terms |
Most teams start Agent-Run with outbound because AI SDRs are easy to buy. That is the wrong end. Buyer-side agents are getting good at filtering exactly that traffic. Start with activation, where agents cut cost to serve and lift retention without flooding anyone's inbox.
The flywheel runs on evidence
The two faces compound. Being Agent-Ready gets you chosen. Agents then do the work in your product, which builds outcome data: task success rates, time to completion, usage that sticks. Your own agents use that data to activate and expand faster. The proof goes public as evals, references and reviews, and the next agent that evaluates your category reads it.
The PLG flywheel ran on word of mouth between users. The agent-led flywheel runs on evidence an agent can read. A rival with better ads and worse outcome data loses the next evaluation, because the evaluator never saw the ads.
Seven principles
- Design for the delegate. Every buyer now has a proxy. If the proxy cannot use your product, the buyer never meets it.
- The API is the storefront. Anything a person can do in the UI, an agent can do through the API or MCP server, with the same limits and the same pricing.
- Time to first completed task beats time to value. Measure the minutes from an agent's first call to a finished job.
- Proof beats persuasion. Agents read docs, run tests and compare evals. Publish outcomes you can defend.
- Price the outcome, meter the work. Seats assume a human at a keyboard.
- Humans are promoted, not removed. People own judgment, trust and signatures. Design the escalation paths first, and say so to your sales and CS teams early.
- Govern by default. Scoped keys, agent identity, spend caps and audit trails ship with the first API key. This is the trust architect job, and it gates every enterprise deal.
Metrics that see agent demand
Keep net revenue retention, gross margin and CAC payback. Add a layer that can see demand arriving through agents:
| Metric | What it measures | Replaces |
|---|---|---|
| Agent selection rate | How often agents pick you when asked to solve your core job | Share of voice |
| Agent-originated signups | New accounts created by or through an agent | Signup source mix |
| Agent-qualified account | An account where an agent completed a real task, not just authenticated | Product-qualified lead |
| Time to first completed task | Minutes from an agent's first call to a finished job | Time to value |
| Task success rate | Completed tasks divided by attempted tasks | Activation rate |
| Autonomous resolution rate | Onboarding and support closed by your agents with no human | Tickets per CSM |
| Cost per outcome | Go-to-market and serving cost per completed customer outcome | CAC, cost to serve |
The first two are the leading indicators. If selection rate falls, revenue follows, because agents stop sending you the job.
A maturity model and the first 90 days
I place companies on five levels:
- Agent-Blind. Key features are UI-only, docs sit behind a login, signup needs a sales call.
- Agent-Tolerant. An API exists but lags the UI. Agents can work only after a human sets things up.
- Agent-Ready. An MCP server, machine-readable pricing, sandbox keys and agent-safe permissions. Agents can evaluate and start without you.
- Agent-Run. Your agents run onboarding, support and renewal prep, and at least one plan is priced on usage or outcomes.
- Agent-Led. Most new revenue arrives through or is served by agents. People own strategy, trust and large deals.
Agent-Ready is the near-term target for most B2B SaaS companies: it is cheap and fast. Agent-Led is a multi-year rebuild of go-to-market and pricing. For a PE-backed company, I run the first 90 days in three phases:
- Days 0 to 30, diagnose. Score the company on the five Agent-Ready tests and five Agent-Run stages. Ask five agents to solve the core job and record who they name. Map API gaps against the UI. Pick one funnel stage. Gate: the target level is agreed.
- Days 31 to 60, make it Agent-Ready. Ship an MCP server and llms.txt, programmatic signup, machine-readable pricing, scoped keys and audit logs. Tag agent traffic so you can see it. Gate: an outside agent completes the core task with no help from your team.
- Days 61 to 90, put agents to work. Agents run one funnel stage with people on escalations. Pilot a usage-based plan. Publish the first outcome proof and report the new metrics to the board.
The second gate is the one that matters. Until an outside agent can finish the core job unaided, Agent-Run work automates a funnel agents are not yet sending traffic to. Treat it like any other bet with a stated kill criterion, the discipline in killing pilots is a capability.
Product-led growth asked whether a user could get to value without talking to you. Agent-led growth asks whether their agent can, and whether yours can carry them the rest of the way. The companies that answer yes to both will be the ones agents choose by default. The rest will find out they lost a deal they never saw.