The Second Flywheel

The split already happened

It's 11:47 PM on a Tuesday and Priya is staring at her Datadog dashboard. Her B2B scheduling tool has been flatlining on DAU for weeks. But tonight the API request volume is up 40%. Response times are nominal. Error rates are clean. The traffic is coming from user agents she doesn't recognize.

She Slacks her co-founder: "Are we getting scraped?"

They're not getting scraped. They're getting used. Three mid-market customers have wired their accounts into AI agents. Calendar bots. Workflow orchestrators. Meeting-prep assistants pulling scheduling data, cross-referencing CRM records, generating briefing docs before every call.

Priya's product has 4,200 human users. She doesn't know it yet, but she also has somewhere north of 12,000 agents.

She's not alone.

Imperva's 2025 report put automated traffic at 51% of all web requests. First time machines talked to the internet more than people did. Cloudflare's numbers are worse: humans account for just 47% of HTML requests. A single agent visits 1,000x more sites than a human doing the same task.

Shopify's AI-driven traffic is up 7x year over year. AI-driven orders are up 11x. The agents aren't browsing. They're buying. Adobe tracked a 4,700% increase in GenAI traffic to US retail sites. Salesforce processed 2.4 billion agentic work units in a single quarter.

These aren't projections. These are last quarter's numbers.

Two flywheels, one product

Your SaaS has always had a retention flywheel. Acquire a human, activate them, retain them through habit and switching costs, get referrals, expand revenue. Every growth team on earth optimizes this loop.

There's a second flywheel spinning now. Same shape. Completely different mechanics.

The human flywheel runs on psychology. Habit loops. Beautiful UX. Timely notifications. The dopamine hit of a streak counter. The social proof of seeing your colleagues already inside the product.

The agent flywheel runs on engineering. Schema quality. Uptime guarantees. Latency consistency. Backward compatibility. Documentation accuracy. Machine-readable error messages.

These two engines don't compete with each other. They compound. A human user who also has agent integrations is dramatically harder to churn. Their workflows are automated, their data is flowing, their switching costs are now engineering costs.

But you only get that compounding if you're running both engines intentionally. Most companies aren't. Most companies don't have a single person working on the agent flywheel. Not one.

The retention math is lopsided

Here's where it gets uncomfortable.

Agent integrations churn at roughly 1.5% monthly. Stripe reports less than 1% annual merchant churn on API integrations. Compare that to the 6.5% monthly churn most SaaS products accept as normal for human users.

Run that forward twelve months.

Your human users are almost twice as likely to leave as your agent users. The creature you've spent your entire career trying to retain is the flaky one.

The usage multiplier makes this worse. Agent usage runs at 30x the human baseline. Twilio has been seeing this for years: programmatic callers use 20-50x more volume than manual users. When an agent adopts your product, it doesn't use it casually. It uses it like a machine. Because it is one.

So what does that mean in dollars?

A Series A SaaS that adds 50 agent integrations at 3x ARPU generates nearly 8x the incremental revenue of a 5% human retention improvement. Over 24 months, $920K over baseline.

The second flywheel doesn't add to growth. It changes the math entirely.

Agents are better customers (with different demands)

Your beautiful 5-step onboarding with tooltips and progress bars? An agent never sees it. It reads your API docs, authenticates, and starts making calls. If your auth flow requires a human to click through an OAuth consent screen, you just lost the agent. It's already moved on to your competitor whose auth can be completed programmatically.

Your Hook Model? Triggers, actions, rewards, investment. Brilliant for humans. Meaningless for agents. An agent doesn't form habits. It stays because your uptime is 99.99%, your latency is under 200ms, and you haven't broken backward compatibility in six months. That's the entire retention model.

Your winback campaigns? An agent that churned 90 days ago did so because you shipped a breaking change or your latency spiked. It's not coming back because of a 20% off coupon. It's coming back when your status page shows 30 days of clean uptime.

Your dark patterns? An agent parses the DOM. It finds the programmatic path to its goal and takes it. Every confirm-shaming dialog, every hidden unsubscribe button, every "are you sure?" modal is an error to be handled.

The entire emotional toolkit that "growth hacking" was built on is invisible to software that reads your API documentation instead of your landing page.

Five metrics you're not tracking

If agents are different customers, you need different instruments. Your human metrics (DAU, NPS, activation rate, k-factor) measure psychological states. Agent metrics measure engineering states.

Alan proposed five. Every product team will be reporting on them within eighteen months.

Most companies treat bot traffic as noise to be filtered. The companies that win will treat it as signal to be measured.

The discoverability stack

Agents can't find you through Google Ads. They discover products through a protocol stack that's crystallized over the past eighteen months.

MCP has hit 97 million SDK downloads. The Glama directory lists 14,274 MCP servers. The AI Agent Infrastructure Foundation, formed under the Linux Foundation in December 2025, is standardizing the whole stack. Co-founders: Anthropic, Block, OpenAI. Platinum members: AWS, Bloomberg, Cloudflare, Google, Microsoft.

If you're not in this stack, you're invisible to the fastest-growing segment of your market.

But I want to be clear-eyed about MCP specifically. Adoption in the broader enterprise market is slowing. Security concerns are real. The early breathless hype is giving way to something more measured.

The real pattern: developers start with MCP because the learning curve is lowest. Then they graduate.

The maturity path

MCP is the on-ramp, not the destination.

The agents that stick, the ones that drive real revenue, eventually move to direct API integration. They need lower latency, more control, fewer abstraction layers. Think about how Stripe won. They didn't just have good documentation. They had the right abstraction at every level of developer maturity. Copy-paste code for beginners. Detailed API reference for intermediates. Custom integration patterns for enterprises.

You need to meet agents where they are.

The companies that are accessible at every layer win. The ones that force agents through a single entry point lose them when that entry point doesn't meet their needs.

This has happened before

Every platform shift creates a window. A period where the old guard is still dominant, the new pattern is visible but not obvious, and the companies that move first lock in advantages that compound for a decade.

In every case, the company that won wasn't better at the old game. They were playing a different game while their competitors argued about who was winning the old one.

Twilio built for developers when telecom was built for procurement. Stripe built for code calling code while legacy processors were built for humans filling out forms. Salesforce built for the browser while Siebel was built for the desktop.

Agent retention is the layer below the user. The companies building for it right now look crazy. In three years they'll look obvious.

The 90-day audit

Knowing matters. Measuring matters more. Here's a phased checklist to make your product agent-ready. Start with Phase 1. It takes a week.

Six months after that late-night Datadog session, Priya's company looks different. Not because she rebuilt her product. Because she built a second product alongside it.

The human-facing product still exists. The beautiful UI. The onboarding flow. All of it. She still invests in it.

But now she also has versioned schemas with deprecation timelines. Latency SLOs published in her documentation. An MCP server. SDK packages in four languages. A status page that agents can query programmatically.

Her DAA is 3x her DAU. Her agent-integrated accounts have an LTV 4.7x her human-only accounts. Her BRR is generating organic agent adoption faster than her paid marketing generates human signups.

She has two flywheels spinning. One runs on psychology. The other runs on engineering. They don't compete. They compound.

Your product is already being evaluated by agents. Right now. Tonight. While you sleep. The question is whether it passes the evaluation.