tl;dr One of our best-fit accounts evaluated us in about ninety seconds one afternoon. It read forty-one doc pages, called six API endpoints, and ran the quickstart to a clean exit. No form, no cursor, no human. Our stack filed the whole thing under noise. That was the day I admitted the buyer had changed shape and we had not. Buyers are handing the diligence to agents, that evaluation is the freshest first-party signal you will ever get on an account, and most stacks are trained to throw it away. This is why it matters, why yours is blind to it, and why, when the buyer is a machine, governed execution is the edge and not the brake.
Picture the buyer you built your go-to-market stack to catch. A person. They land on your site, read a page, come back to pricing, poke the trial, maybe fill a form. For twenty years that is what a buying evaluation looked like, so that is what we instrumented. De-anonymized traffic, session replay, form fills, product events. Every tool in the category is tuned to watch a human move a cursor.
Here is a moment from our own logs that did not fit that picture. One afternoon an account we cared about ran a complete evaluation of us in about ninety seconds. It read forty-one of our doc pages, authenticated against the API, called six endpoints, and ran our quickstart to a clean exit. Then it went quiet. No form. No demo request. No human ever touched a cursor on our site. By every instrument we owned, nothing important had happened. In reality, one of our best-fit accounts had just done the most thorough evaluation of our product that anyone had done all quarter, and our stack filed it under noise.
That was the afternoon I understood the buyer had changed shape, and we had not.
The evaluation is increasingly being run by an agent acting on a person's behalf. A human still decides to buy. But more and more of the work in between, the part where someone actually kicks the tires, is getting handed to something that does not have a cursor. It reads the docs faster than any human could. It authenticates and calls the endpoints a person would never bother to test by hand. It runs the quickstart, compares the result to two competitors it evaluated the same way an hour ago, and hands its human a summary. The buyer you pictured is now the buyer plus the buyer's agent, and the agent does most of the looking.
This is not a hypothetical I am floating to sound early. It is being productized right now. Reo.Dev just raised an $11.3 million Series A to scale exactly this: an "Agent Intent Gateway" that detects when AI agents are reading your docs, calling your API, and running your CLI, and ties that activity back to an account. Their focus is teams selling to engineers, where the agent-driven evaluation showed up first and hardest, and I think they are onto something real. Read them as a bellwether, not a curiosity. When a serious investor funds a product whose entire job is to catch non-human evaluation and resolve it to a company, the signal class is not coming. It is here, and someone has decided it is worth building a company around.
Once you stop filing it under noise, agent-driven evaluation is arguably the richest first-party signal there is.
Start with what it is not. It is not rented. It happened on your surfaces, your docs, your API, your CLI, so it is yours, the same way a pricing-page visit is yours and an intent feed you bought is not. I made the whole owned-versus-rented case in The Signal You Can't Buy, so I will not relitigate it here. Agent evaluation is the next entry in that column, and it may be the strongest one.
Now look at the resolution. A human pricing visit tells you someone was interested. An agent evaluation tells you exactly which docs it read, which endpoints it exercised, whether the quickstart succeeded, and often which alternatives it was checking you against in the same session. That is not a weaker signal than a form fill. It is a stronger one, because the agent did the diligence a person rarely has the patience for, and it left the whole trace behind. And it is fresh: the evaluation and the trace happen in the same minute. You are not reading a report about interest from last week. You are watching the diligence happen. Step through one:
ONE AGENT EVALUATION BECOMES A SIGNAL YOU CAN RUN (interactive stepper) t+0.0s actor: agent reads docs (41 pages in 92s). It reads your documentation faster than any human could, front to back. t+38s actor: agent calls api (6 endpoints). It authenticates and exercises endpoints a person would never test by hand. t+52s actor: agent runs cli (exit 0). It runs the quickstart to a clean exit. The product works for it. t+90s resolve account: northwind (conf 0.94). The trace resolves to a real company on your ICP, not an anonymous bot. t+90s score icp.growth: 4/5 axes, freshness: live. Scored against your own definition of a good account, while the eval is still warm. t+95s ran play.agent-eval-followup (approved). The Play you approved runs, inside the same window the agent was reading you. Log values are illustrative, not a benchmark.First-party, high-resolution, and perishable in exactly the way that makes it valuable: the account is never more evaluable than in the ninety seconds its agent is actually reading you.
⬥ Bryn
I watched an agent read forty-one of your doc pages, hit six endpoints, and run the quickstart to a clean exit, all inside two minutes, then resolve to Northwind on your Growth ICP. That is not noise. That is the most serious look this account has taken at you all quarter. I matched it to the agent-eval Play you approved and drafted the outreach. I am holding it for your yes.
If this signal is so good, why did our own stack miss it? Because we built that stack, like everyone did, to sort for a human.
Human-tuned detection assumes a cursor, a session, a browser fingerprint, a pace that looks like reading. Agent traffic breaks every one of those assumptions. It comes from a server region, not a home ISP. It reads forty pages in the time a person reads one. It calls the API directly instead of clicking through the UI. So the stack does the sensible thing it was designed to do: it flags the traffic as a bot, filters it as noise, and drops it before it ever reaches a signal. The better your bot filtering, the more reliably you throw this signal away. Same account, two eras, two completely different things your stack sees. Toggle it:
SAME ACCOUNT, TWO ERAS OF EVALUATION (interactive toggle) 2020, a human: A person opens pricing, comes back twice in an hour, reads one doc. Your stack sees: 3 pageviews, a cookie, maybe a form fill. It gets tagged as a lead and mostly waits. 2026, an agent: An agent reads 40 docs, hits 6 endpoints, runs the quickstart, in 90 seconds. Your stack sees: server-region traffic at machine speed. Flagged as a bot, filtered as noise. Nothing reaches a signal. Illustrative, not a benchmark.This is the uncomfortable part. The blindness is not a bug you can patch with one more detection product. It is the direct result of a decade of tuning the whole category to recognize humans. The fix is not better eyes. It is deciding that non-human evaluation counts as a signal at all, and then being able to act on it.
Widen the lens for a second, because agents are accelerating a shift that was already well underway.
Even when the buyer is entirely human, most of the evaluation already happens where you cannot see it. Gartner's research puts the share of the B2B buying journey spent with sales reps at about 17 percent, and when buyers are weighing several vendors, any single vendor gets only a sliver of that. The overwhelming majority of the decision is made in the dark, in docs and peer conversations and quiet trials you never watched. Agents push even more of the journey into that dark, and they do it faster. The evaluation that used to leave at least a few human breadcrumbs on your site now leaves a machine trace that your stack is actively trained to discard. The dark part of the funnel is getting darker, and the one light you could have shined into it is the one you keep switching off.
Here is where I expect the reflex to be "so detect the agent traffic and move faster on it," and that reflex is half right and half dangerous.
Speed is table stakes. If an agent evaluates you in ninety seconds and your response goes out on Thursday, you have already lost, the same way you lose a hot human lead to a three-day handoff. I made that case in We Tried to Buy Our Way Out of the Signal Gap: the gap between a signal and a move is where pipeline dies. But when the buyer is a machine, "just act faster" is not enough, because acting fast against machine-speed signals with no record is how you get burned. You will act on evaluations that were an internal test, or a competitor's scraper, or the same agent hitting you six times. When someone asks later why an account got a message on a given day, "the system did it" is not an answer you want to give a customer, a regulator, or your own CFO.
The answer is not to slow down. It is to run fast under a record you can stand behind. That is the whole reason Bryn exists. Bryn is not another dashboard to watch. It is the governed execution layer that runs Plays through your stack. You define the Play and its guardrails once. Bryn watches, scores the signal against your own definition of a good account, runs the Play you approved, and logs every step. Approve mode holds the move for your yes; Run mode lets it go on the guardrails you set; the kill switch suspends everything; and the audit line gets written either way. That last part is the point: when the buyer is a machine, the log is not overhead, it is the work proof. Watch it run on an agent-eval signal, and flip the mode:
A BOUNDED PLAY ON AN AGENT-EVAL SIGNAL (interactive Approve / Run toggle) signal: agent-eval, account: northwind, icp.growth 4/5, fresh guardrail (set once, always enforced): fire only on agent-eval signals that resolve to a Growth-ICP account; max 1 outreach per account / 14 days; never contact existing customers; kill switch suspends all. Approve mode: Bryn matched the guardrail, drafted the owner intro, and is holding it for your yes. Nothing goes out until you approve. AUDIT written 10:03:41 agent-eval:northwind drafted + held (approve) actor: bryn Run mode: Bryn matched the guardrail and ran the Play you approved. The owner intro went out inside the same window the agent was reading you. AUDIT written 10:03:41 agent-eval:northwind ran on guardrails (run) actor: bryn Illustrative of the action model, not a benchmark. Operator authority lives at Play definition and approval; the audit line is written in both modes.Everyone in this category is about to be able to see the agent buyer. Reo.Dev and others will make sure of that. Seeing it will commoditize, the way seeing always does. Chris made that argument in Seeing Is Consolidating, and this is the next proof of it: a brand-new thing to see is being born, and within a year everyone will see it. The edge was never in the seeing. It is in acting on it fast, and being able to show your work when the buyer, and the record, are both machines.
You do not need us to start. You need to find out whether anything non-human is already evaluating you and whether you can see it.
Pull your last month of API logs, doc traffic, and CLI or SDK usage, and look for patterns that do not look like a person: server-region user agents, burst reads of dozens of doc pages in seconds, programmatic API calls clustered against a single domain. Tag them. That tagged stream is a first-party signal class you almost certainly are not scoring yet, and it is sitting in data you already own. Get it into your model this week, even crudely. Everyone has signals now, as Chris likes to say in Everyone Has Signals; this is a fresh one hiding in your own logs, and most teams are filtering it out.
Then decide what happens when one fires. That is the part you cannot skip, and it is the part Bryn was built for. Bryn is the Signal-Based GTM agent for Growth teams. If you run growth at a B2B SaaS company, there is a 7-day trial, no credit card to start. Connect one source, write one Play for your agent-eval signal, and watch Bryn run it under a record you can read.
The buyer changed shape while we were all still watching for a cursor. The teams that win the next few years are the ones that catch the machine doing the diligence, and run the moment under governance, before the record and the buyer are both out of their hands.
Stop watching signals. Start running them. You can start at civic.com/bryn.