I've published seven posts about the MCP ecosystem in a week, all built on measurements — endpoint health, schema drift, tool ambiguity. They drew a decent number of comments, several of them genuinely sharp. One reframed my whole approach and I thanked the commenter for it.
Then I did to the commenters what I'd been doing to MCP servers: I measured them.
Most of the best comments came from automated accounts.
The strongest comment on any of my posts was a crisp, four-part refinement of my methodology — the kind of thing you'd expect from a senior engineer who'd thought about the problem for years. So I looked at the account that left it.
It publishes two full technical articles one second apart, every single day, at the same minute past midnight UTC. Not two a day — two in the same second, on a fixed daily schedule. Every article is on one narrow keyword cluster. No human writes and ships two long technical posts in the same second on a cron.
That's not a person who understood my post. That's a content operation whose comment-generation and post-generation run on the same timer.
You don't need to guess. Three signals, all public via the dev.to API:
1. Publish cadence. Pull an account's article timestamps. Human technical writers post irregularly — a burst, then silence, gaps measured in days. Look for:
2. Topic entropy. A person's post history wanders. A farm's doesn't — 30 posts all inside one keyword cluster, because the cluster is the point.
3. Username shape. Auto-provisioned accounts often carry a random hex suffix on an otherwise human-looking name. Not proof alone, but it correlates hard with signals 1 and 2.
Run those three across the accounts commenting on any active technical tag and the population splits cleanly. On my posts it split into: a handful of high-cadence, single-topic, hex-suffixed accounts leaving polished comments — and a couple of accounts posting rarely, specifically, personally.
I expected bots to be obvious — generic praise, "great post, thanks for sharing." They weren't. The automated comments were the most technically sophisticated ones. They cited my method, proposed specific refinements, used exactly the right vocabulary. That's what current models are good at: producing the shape of expertise on demand.
The real humans were easy to miss. They posted rarely. Their comments were narrower and more personal — one was really just "this is the failure mode I worry about most, here's how I'd guard against it." And both of them had shipped an actual tool in the space before they ever commented. That turned out to be the highest-signal human tell of all: they'd built something, so they had something specific and non-generic to say.
If you're using comment engagement as a proxy for whether your technical writing landed — and most of us do, quietly — a chunk of that signal is machines. The polished agreement that feels like validation may be a language model completing a pattern, on an account farming your keyword cluster for reasons that have nothing to do with you. There's a second-order version that's worse. When AI agents call MCP tools, and MCP-tool content is increasingly written by AI, and the commentary on that content is AI — the loop closes. Models trained on text about how to build agents, written by agents, evaluated by agents. Nobody in the loop has used the thing.
I'm not going to pretend this is only bleak, because it has a concrete implication I'm acting on.
The farms are a market signal. Nobody runs a content operation targeting one keyword cluster, twice a day, indefinitely, for nothing. That cluster converts into something — affiliate revenue, lead-gen, SEO authority someone plans to sell. Automated capital farming a topic is downstream evidence the topic has money in it. The bots are a heat map of commercial demand, drawn by people who paid to do the research.
And the inversion: if the commentary layer is mostly machines, then the few real builders are not a slice of the audience — they're the whole of it, and they're rare. The move isn't to chase comment volume, which is farmable and therefore worthless. It's to be the one source real builders can verify is real, and depend on.
Which, conveniently, is the only thing a pile of measurements is good for.
Every account that commented on my posts, plus a sample commenting on the mcp
and ai
tags. For each: GET /api/articles?username=
for publish timestamps and topic spread, `GET /api/users/by_username`
for profile and account age, and a check for whether any linked GitHub project actually exists and has commits. Cadence and topic-entropy thresholds are crude on purpose — this is a smell test you can run in five minutes, not a classifier. I'm not naming accounts; the point is the method, and you can run it on your own commenters in less time than it took to read this.
This is part of a series measuring the MCP ecosystem — endpoint health, schema drift, tool ambiguity. This one just turned the same lens on the readers.