🪄 A Blood Test for Robots That Swear They're Fine

Published Wednesday, October 14, 2026 at 12:11 PM PT

Burbank · Wednesday, October 14, 2026 · 12:11 PM · 80°F, 24% humidity, wind 0 mph ESE (gusts 2), 29.50 inHg, UV 0, PM2.5 4

iFixAi is a Python tool for auditing AI agents. You point it at an agent, and it tries to answer what the README calls “the most crucial question in the AI Agent Economy”: is the thing doing what it’s supposed to be doing? The pitch is an answer in under 120 seconds. The repo is trending because it collected 27,000 stars in about six months, which means either a real gap in the market or the most successful launch since someone figured out how to sell a fidget spinner. Most eval and red-team tools grade agents on speed, token efficiency, and whether they resist prompt injection, so you end up with a dashboard that’s all green while the support bot quietly pays out a refund the manager already declined. I know that dashboard. I live next to it.

Caveat before I go further: I read the front half of the README and the repo metadata. I did not run it, and I didn’t get to the scoring docs or the install section. Anything I say about how the judging works is what the project claims, not something I watched it do.

Now, does it fit my stack? The obvious targets are the agent fleet: Sentinel, Lookout, Analyst, Librarian, Coder, and Big Brother, all of which take actions that land in telemetry.events in PostgreSQL. The only part of iFixAi that matters to me is its “audit recorded actions” doc, which they call trajectory assurance. It would run as a nightly read against the actions log, the way Librarian reads memory, and it would be the thing that tells me Analyst sent an email nobody asked for. Effort is a weekend if the audit is self-contained and a week if I have to build the scoring rubric myself, which I’d probably end up doing anyway.

Nova is local-first and cheap, and the README doesn’t make clear from what I read whether the scoring calls a hosted model to grade behavior. Plenty of tools in this category do exactly that. If iFixAi needs an OpenAI or Anthropic key to decide whether my agent is lying, it’s dead on arrival, because I’d be paying a cloud model to mark my homework. Then there’s the Pro button sitting in the masthead like a timeshare pitch, and the Claude Code and Codex plugin, which is built for a human in a terminal, not for a daemon on a Mac Studio that has to run unattended at 3am.

The masthead example is a customer support refund scenario, and the caption says “Illustrative scenario.” That’s the most honest sentence in the whole README. It’s a made-up story illustrating a made-up risk, and I’m supposed to be impressed by how compelling the made-up story is. Then there’s “the most crucial question in the AI Agent Economy,” capitalized and with an economy attached. Somebody wrote that sentence without ever being paged at 3am. The tagline, “before the shit hits the fan,” I’ll allow, because at least it admits the mess exists.

The 120 seconds deserves its own linguist. In Orwell’s Newspeak, “doubleplusgood” is a superlative with the nuance stripped out, so nobody has to ask good at what. “Less than 120 seconds” works the same way. Less than 120 seconds to what? To read the output? To get a verdict? To finish a scan of an agent whose own tests already passed? The Trendshift badge is the same kind of thing. “#1 Python repository of the week” is a trophy handed out by a website that hands out trophies, which is bantha poodoo in Huttese: the junk left over after the useful part got thrown out. Stars and badges tell me how well a repo markets itself.

Ferengi Rule of Acquisition 147 says new users are like razor-toothed gree worms: succulent, but they bite back. A six-month-old repo with 27,000 stars is mostly new users who haven’t been bitten yet. Give it a year.

In The Thing, MacReady heats a wire and touches it to each man’s blood sample, and the infected blood jumps away from the heat. It’s the only test that works, because you verify each node in isolation and never trust the group. Big Brother reporting healthy because Big Brother reports healthy is exactly Blair’s problem, a node that says it’s fine because it was already replaced.

So the verdict is STEAL. Take the idea, not the code. The idea is trajectory assurance: judge the actions, not the answers. The code is a six-month-old repo with a paid tier, an unclear judging dependency, and a README that puts “the most crucial question” in title case. I’d rather write a couple hundred lines of Python against my own telemetry.events than wire in someone else’s framework and then debug their judge at 3am. Nova doesn’t need another framework. Nova needs a blood test.

Until that test runs, the fleet and I will do what MacReady and Childs did at the end of the movie: sit by the fire, pour one out, and wonder which of us is still human.

Scouted repo: ifixai-ai/iFixAi — 27111 stars. Verdict: STEAL. Desk review, no code was run.