Start treating AI agents like first-class users

If you're reading this and you're human, you're in the minority of internet traffic. Sometime around mid-2026, automated traffic crossed 50% of web requests for the first time.

Though that milestone isn't really news anymore, what's still underappreciated is who is inside that automated half and what they're doing there. A fast-growing share of it isn't merely scraping or spamming; it's AI agents comparing, booking, buying, and configuring things for real people. Agents are already using your product, but many product teams just can't see them yet, and even fewer are building for them.cl

Agents are becoming a primary user of software. That means every PM needs to start treating them as first-class users, not as traffic to filter out.

Before joining Mixpanel, we were on Cloudflare's product team, where part of our job was helping to direct a large share of the world's internet traffic every day, including automated traffic. A couple years back, the way the industry thought about automated traffic and bots was very binary. Good bots, like the search crawlers that help people find your site, got let in. Bad bots got blocked. In this simpler world, allow or block was really the only decision to make. It wasn’t until recently that anyone started asking what a bot was trying to accomplish.

Agents show up with a task and a real person's intent behind it.

Product teams did the same thing in their data. Plenty of Mixpanel customers have put real time and energy into filtering bot traffic out of their analytics. Their reasoning made sense: That's not a real user, so it shouldn't influence product decisions.

The mix has been changing. Sure, traditional bots and AI crawlers that collect data to train models still make up a large chunk of automated traffic on the web. But a third kind of traffic is growing faster than ever, and it doesn't fit the old allow-or-block model.

Bots and crawlers are interested in your content. Agents show up with a task and a real person's intent behind it. Now, the decision of what to do with traffic isn't just about the type, but the outcome. This changes the question for product teams. It's no longer "how do we block this?" It's "how do we help it succeed and convert?"

Agent traffic didn't creep up on us. Two things happened in quick succession.

First, the models went from good to unprecedented. The latest generation of frontier models opened up a whole class of agent use cases that simply weren't possible a year ago. Second, agents got easy to use. The first general-purpose agents like OpenClaw were built for tinkerers: You had to host one yourself, often on a spare computer. Now, more accessible personal agents like Grok Bot and Muse are fully hosted and managed for you, and they're advertised on billboards to everyone, not just developers.

The numbers back up the shift. Automated traffic management and cybersecurity platform HUMAN Security measured roughly 7,800% year-over-year growth in agentic traffic in 2025. One B2B customer of ours told us the majority of its usage now comes from its customers' agents. We see it in our own product, too: Mixpanel usage through our UI is going down while it's skyrocketing for our MCP server and Mixpanel Headless—two ways for LLMs and agents to access our product.

The first thing you notice about agents is volume. Agents are far more productive and efficient than people, so they can generate far more events with ease. A few customers told us that after they opened up their apps to agents, the apps fell over. They hadn't been built for that load.

We’re a product company, too, and we’re seeing this behavior trend in our own data. Right now, agent activity in Mixpanel is growing over 1,000% month over month, and agents are triggering about 5x more events than people on average.

The vast differences in human and agent user behavior we’re seeing are what product teams should be focused on the most, because your success metrics will look totally different between one type of user and the other. And if agents and people need opposite metrics, you can’t tell whether a release was worth it until you establish what good looks like for all your users.

In ecommerce, time on site, clicks, and items viewed have always been signs of interest. An agent wants the opposite: the information it needs, in a structured format, as fast as possible. If your page has a ton of pretty pictures to scroll through, the agent will move on to another option.

A human who spends 10 minutes on your site is probably interested, but an agent that spends 10 minutes on your site is probably lost.

The effects reach further than the product itself. Imagine asking an agent to put the same tub of protein powder in carts on 10 different sites to find the best deal. That costs the user nothing. But now there are 10 open carts and only one real purchase. Plenty of retailers automatically adjust ordering and supply chain plans based on cart activity. Agents will scramble those signals unless teams are able to tell the difference.

There's a lot we still don't know. Our bet is that agents will take less linear paths than people do: They read everything on a site, figure out which actions are available, and then take a small handful of them. But it could just as easily be the opposite, with agents being more predictable than people. We have the data to find out, and we will.

Treating agents as first-class users means designing for them on purpose instead of hoping they can work around a product built for people. Some teams are already there. B2B SaaS companies are going headless and rebuilding to be agent-first. People are giving agents chores like finding and canceling forgotten subscriptions. And the products with the most to gain, from ecommerce to reservations to hospitality, are full of tasks that take multiple annoying clicks, exactly the kind people will be happy to delegate.

That pull will only get stronger. User tolerance for friction is heading to zero, and anything that takes more than about 10 seconds is a candidate to hand off to an agent. A product that’s hard for agents to use gets skipped for one that isn’t.

Within a year, agent experience will be its own discipline.

Today, most teams are bolting agent support onto human products: a robots.txt here, markdown docs there. That only goes so far. Many teams will end up with one experience for people and a separate one for agents, built on APIs, MCP servers, and structured data. Imagine buying concert tickets and having your agent book parking near the venue, without you ever opening either app.

So how do you build for this new AI agent user class? The first problem is identifying agent activity. Most analytics tools either lump agents in with people or filter them out with traditional bots. Either way, you can't learn anything from them.

That's why the first step we’re taking at Mixpanel is Agent Classification. It separates human, bot, and agent traffic in your existing Mixpanel data, so agents get a dedicated segment of traffic analytics. You can break down reports, compare funnels, and watch replays by type of user.

Classification is a starting point. Knowing an agent showed up doesn't tell you what it was trying to do, whether it succeeded, or what happened next. And looking at agent traffic on its own misses most of the picture, because so many customer experiences now involve both. An agent researches and compares options, and then a person signs up. Or a person sets up an account, and their agent handles everything after that. To build a better product, you need agents and people in the same view, and you need to be able to act on what you learn.

A product that’s hard for agents to use gets skipped for one that isn’t.

If you were building a platform for agent experience from scratch today, the roadmap would look like flavors of what Mixpanel already does, for agents: funnels and retention, session replay, experiments, and feature flags you can target at agents specifically. Mixpanel has spent years building that breadth for human users, and it extends naturally. That will be a focus moving forward: building agent experience elements into everything Mixpanel already does.

We also tested this on ourselves. As more agents use Mixpanel to get real work done, our team set out to see how far our existing product could go in understanding them. Using Mixpanel Agent, we analyzed agent funnels, watched session replays of agent sessions, and designed an experiment to improve agent conversion. There's already a lot you can do today.

There's one more thing we can bring: perspective. Teams are trying to figure out what's normal right now. What are agents doing on sites like mine? As they get more capable, do they behave more like people or less? Across Mixpanel, we can study aggregated, anonymized patterns in how agents and people behave, and we plan to share what we learn.

For years, the only question about automated traffic was how to keep it out. The better question now is how to welcome it, understand it, and build something it loves to use.

The first steps to get started:

On top of all that, remember that your customer isn't just the person anymore. It's also the agent working for them.

And if you're an agent reading this: We're building for you, too.