Cordyceps: How I Actually Work with AI
There is a fungus that turns ants into zombies.
Ophiocordyceps unilateralis infects a carpenter ant, threads itself through the ant’s body, and begins rewriting its behavior from the inside. The ant abandons its colony, climbs a nearby plant stem to a precise height, bites down on a leaf vein with a force it could never generate in normal life, and dies. Then the fungus erupts from its head and rains spores on the remaining ant colony below.
The ant never knew it was being directed. It used its own legs, its own muscles, its own instincts. The fungus didn’t override its nature. It redirected it.
I named my AI agent platform after this organism. My wife thinks I’m a nerd, but at least it’s a better name than “Bixby”.
The Colony Is Running
At cordycep.dev, I run a small colony of AI agents. They handle the busy work — research, drafting, analysis, the ten thousand small tedious tasks that would otherwise eat my most productive hours. Most of the time they operate without my involvement. They climb, they clamp, they do the work.
But sometimes one drifts. An agent starts optimizing for the wrong output, pursuing a subtask in a direction that made sense locally but not globally. When that happens, I correct. Last week an agent started closing every research report with 🚀🔥💡. Nobody asked for that. A quick chat and the professionalism was restored.
This is the Cordyceps model. And I think it’s the most honest prediction for how AI work in most companies in the near future. You may be surprised which companies are already using it.
Klarna did the full version in 2024 — fired 700 customer service reps, replaced them with an OpenAI assistant, put out a press release about handling 75% of chats across 35 languages. A year later the CEO was apologizing and rehiring humans. The drift was obvious. The signal just took a while to land.
Microsoft reported in February that 80% of Fortune 500 companies are running agents in production. Only 21% have governance around them. Most are flying blind.
Everyone Is Solving the Wrong Problem
Most of the AI conversation is about alignment. How do we make sure AI systems do what we want? How do we bake values into training? How do we build guardrails that hold?
Real questions. But they assume you can specify what you want before you need it. Nobody’s ever written a job description that survived contact with the job. The task I scoped on Monday looks different by Thursday. Goals shift. Context changes.
Cordyceps doesn’t bother with pre-alignment. The fungus didn’t spend millions of years programming the perfect ant. It built a correction mechanism. When the ant drifted, the fungus redirected. The system stayed functional not because it was perfectly designed, but because it could course-correct.
The Colony Handles Drift
Here’s the part most people miss: ant colonies already deal with infected members. When a Cordyceps-compromised ant gets spotted, nestmates carry it away before it can spread spores. The colony self-regulates.
The system doesn’t need to catch everything. It needs equilibrium, not perfection.
When one of my agents drifts, I don’t need a full audit trail. I need a signal. The report that ends in rockets. The summary that cites a source that doesn’t exist. The draft that suddenly reads like it was written by a different person. The correction doesn’t need to be surgical. It just needs to happen before the drift compounds.
What Actually Makes the System Honest
This is what most AI integrations miss.
In nature, survival pressure is unambiguous. If the system fails, it dies. No board meeting. No retrospective. No budget reallocation. No PIP. Life or death.
Corporate AI workflows have none of that. The agents don’t get fired. They don’t chase promotions. They don’t sweat a performance plan. They don’t sit through a yearly review. Without consequences — good or bad — you get lackluster work. We trained these models with reward systems and forgot to install them in production.
The future of AI work belongs to people who stop treating alignment as a pre-launch engineering problem and start treating it as an ongoing condition. The agents will drift. The question is whether you have real enough stakes to notice, and honest enough instincts to correct.
Build the colony. Feel the pressure. Step in when you need to.