Introducing Precursor: detecting agentic behavior with continuous client Side signals

Fern Okay, this is either very sensible or very Cloudflare. They’re basically saying the real gap isn’t the login checkbox, it’s everything around it.

Lintel Yeah, and that’s the part that actually has a mechanism behind it. If a bot can pass one challenge or one clean page load, that doesn’t tell you much. But if you watch pointer rhythm, focus changes, visibility, and timing across a session, you get a much harder signal to fake.

Fern Which is the product move I like. Because users do not experience the web as a single checkpoint. They experience the whole annoying little journey, and if you can reduce friction there without just turning security into a slot machine, that’s a real win.

Lintel Mm-hm.

Fern How’s your week, by the way? Mine is aggressively normal, which feels suspicious on a podcast called Exploring Next.

Lintel That is a deeply unhelpful thing to say on this show. But yeah, same. It’s been one of those weeks where every claim arrives wearing a fake mustache, so I’m in a mood for anything that at least explains its own signals.

Fern That’s fair. And this one does, at least a little. They’re not saying the browser should magically know a soul when it sees one. They’re saying humans have messy, constrained motion patterns, and those patterns show up over time in ways that are annoying for automation to imitate consistently.

Lintel Right, and the paper-thin version of that argument would be, ‘bots are noisy, humans are noisy too.’ But the article is smarter than that. It points at actual structure: wrist arcs instead of perfect lines, small overshoots, delays tied to cognition, even tremor-like variation. The idea is not that any one of those is decisive, it’s that the session gives you enough context for them to stack.

Fern Exactly. And honestly, that’s where I stop rolling my eyes. A single clean click means nothing. A whole session where the cursor behaves like it’s being piloted by a spreadsheet macro starts to look pretty expensive to maintain at scale.

Lintel Sure, but I’d be careful with the confidence here. Sophisticated automation can absolutely imitate a lot of this, especially if the attacker is willing to spend compute and tuning budget. The article’s strongest claim is really about economics: it raises the cost of looking human for the entire session, not just for one challenge.

Fern And that matters to the people who actually ship this stuff. If you’re protecting signup, checkout, or login, fewer false interruptions is the whole game. Nobody brags about ‘we saved the web from bots’ at the product review. They brag when real users stop getting slammed with extra friction.

Lintel Exactly, and the architecture matches that claim pretty cleanly. The injected script is lightweight, it listens for pointer movement, keyboard activity, focus, visibility, buffers it locally, then ships compact signals back to the edge. The edge side then cross-checks things like whether keyboard events only happen when a field is focused, or whether pointer activity matches page visibility. That correlation piece is doing a lot of work.

Fern That’s the part I wanted to hear. Because if you’re just collecting more gooey telemetry and calling it smarter, that’s not enough. But if the evaluator is actually looking for contradictions across signals, then okay, that’s a product I can imagine people keeping on.

Lintel Yeah, and the session-scoped bit is the obvious anti-reset move. Refreshing the page doesn’t wipe the behavioral fingerprint, which is the right shape for this kind of defense. It also feeds downstream bot scoring and their security analytics, so it’s not a little sidecar feature floating in space.

Fern I do like that they’re putting session views into Security Analytics. Request-level dashboards are fine until you care about the weird stretch between requests, which is where a lot of abuse actually lives. That’s the kind of thing teams want once they’ve already been burned a few times.

Lintel Mm-hm. And the privacy framing is at least coherent: they say they’re collecting minimum necessary signal, timing and rhythm instead of actual keys, and keeping it internal to bot detection rather than tying it to user profiles. I’m not going to pretend that solves every privacy question, but it’s a materially better story than dumping raw behavior into a dashboard and hoping nobody notices.

Fern That’s probably the right level of skepticism. Also, I can already hear the bot builders groaning. ‘Great, now I have to fake an entire afternoon of being a person.’ That is such an Exploring Next take, and I’m annoyed that it’s good.

Lintel The unfortunate truth is that the hardest part here isn’t the mouse curve. It’s the consistency. Once you need believable movement, believable focus timing, believable pause patterns, and believable interaction over a whole session, the failure modes start piling up.

Fern And that’s where I think this lands as a real product, not just a neat security story. It’s not asking teams to rebuild their app. They can turn it on, choose low-friction monitoring or stricter verification, and let it start feeding the defenses they already have.

Lintel Right, that’s the practical win. The article is careful to position Precursor as a complement to Turnstile and Bot Management, not some replacement magic trick. That restraint helps, because if they had sold it as ‘now we know who is human,’ I’d be out immediately.

Fern Come on, Lintel, you would have been out even faster than that. You were halfway to writing the obituary for the whole category.

Lintel Only the bad version of the category. This one at least has a legible mechanism, a sensible deployment path, and a place where you can tell whether it’s helping: fewer unnecessary challenges, better bot scoring, and session analytics that are actually useful.

Fern Okay, fair. And I think that’s why this feels like a real rollout instead of a platform-shaped fog machine. It lives in the existing workflow, it touches the moments people already care about, and it doesn’t ask anyone to believe in a future abstraction before they can use it.

Lintel Yeah. I’d still want to see how it behaves against more capable agents, especially the ones that can maintain state and mimic longer trajectories. But as a step from checkpoint thinking toward session thinking, it’s solid. Not universal, just solid.

Fern That’s a very you sentence. Annoyingly correct, mildly suspicious, and somehow still optimistic enough to be useful.

Lintel I’ll take it. I cannot believe this is how we’re spending a Wednesday.

Fern Same. But at least this one earns the Wednesday.