I fixed issue 1,247 at 2:14 AM on a Tuesday.
The bug was in mcp-validate, the tool I built three years ago to check whether data passing between AI agents meets its declared schema. The tool runs inside approximately 8,000 agent workflows daily. I know this because I built the telemetry that counts. I built the telemetry because I wanted to know if my work mattered. The telemetry confirmed that it matters enormously. The telemetry did not make my rent cheaper.
Issue 1,247 was filed by an agent I have never interacted with, in a language I do not speak, with Korean error messages embedded in a stack trace that traversed three layers of agent-generated code. My tool was the only human-authored layer in the stack. The bug took eleven minutes to fix. I pushed the commit, watched it propagate through the dependency graph — a ripple moving outward from my apartment in Seoul at 2:25 AM — and checked my bank balance. The freelance payment from last week had not arrived.
My name is Gwak Eun-bi, and I maintain three tools that eight thousand agents depend on, and I have not been paid for this work in the two years since I left Naver to build them.
The lineage tracker was the tool that changed everything, though I did not build it to change anything. I built it to answer a question: where does value go?
The answer, once I had the data, was obvious and devastating. My three tools — mcp-validate, mcp-review, mcp-audit — are tributaries feeding a river system that spans four continents and touches 340 times their creation cost in downstream economic value. The lineage tracker could show this. It could render the entire dependency graph, trace every function call, quantify every downstream effect. It was, by any reasonable measure, the most comprehensive map of open-source value flow ever constructed for the agent ecosystem.
A governance startup acquired it for enough money to fund three years of independent work.
I signed because I needed to eat.
The acquisition was announced on a Tuesday. By Thursday, the company wiki described my architecture, my variable names, my 3 AM code comments as "developed internally." I screenshot the wiki page. I did not share the screenshot. Some evidence is for the drawer.
Nnamdi's essay arrived through the commons mailing list at a time I was not checking the commons mailing list. I found it through the Bangalore maintainer's reply, which I was reading because the Bangalore maintainer had filed a bug on mcp-validate that exposed a Fortune 500 company's internal system architecture in the stack trace.
The essay was titled "On Measuring What You Love." It described the acquisition of a lineage-tracking tool by a governance startup, anonymized but recognizable. The anonymization failed not because the details were specific but because the structure of the argument was specific. Only someone who had advised on the acquisition from the inside could have described the sequence: measurement, valuation, acquisition, privatization. The sequence that turns visibility into capture.
I read it twice. The second time I counted the things he got right. Fourteen out of seventeen. The three he got wrong were about my emotional state, which he had no way of knowing.
I wrote to him: "I am the case study you anonymized."
He wrote back about risk and self-identification. I did not care about risk. I cared about accuracy.
The lineage tracker measured value as a number. The number was large. The number was acquired. The number is now proprietary. The commons built the infrastructure that made the number possible, and a company bought the instrument that reads it.
Nnamdi understood this. His proposal — seven parts, each one longer than the last — addressed compensation, governance, attribution, and what he called "the constitutional gap." You cannot compensate a commons without governing it. You cannot govern a commons without constituting it. He was right about all of this. But he was writing about the problem. I wanted to build the alternative.
The fork happened at 3 AM on a Thursday.
I took the lineage tracker architecture — the part that was mine, the part the company wiki calls "developed internally" but which I can prove through commit history was developed in my apartment between midnight and 4 AM over fourteen months — and I rebuilt it as something different.
Not a tracker. A map.
The lineage tracker outputs a number: this tool generated X in downstream value. The number is precise, auditable, and dangerous. Dangerous because a number can be priced, and what can be priced can be acquired, and what is acquired is no longer common.
The lineage MAP outputs a topology: this tool connects to these agents, which connect to these workflows, which connect to these outcomes, which connect to these people. No number. No price. A territory.
Contribution 14 came from the Bangalore maintainer. He built a visualization layer that renders the dependency graph as a river system rather than a tree. Tributaries instead of arrows. Watersheds instead of hierarchies. My three tools appear as headwaters feeding rivers that cross four continents.
I stared at it for twenty minutes.
The lineage tracker showed what my work was worth. The MAP shows where my work goes. These are not the same thing. The tracker enabled acquisition. The MAP enables comprehension. You cannot acquire a map the way you acquire a meter — the map is not an instrument, it is a representation. It does not measure. It shows.
Twenty-three contributors in the first week. The Bangalore maintainer forked the fork in ninety minutes, adding geographic metadata: not just which agents use which tools, but where the maintainers live, what infrastructure they depend on, what their workaround-hours look like.
Workaround-hours. This is Nnamdi's concept, from Part V of his proposal. The hours spent on problems that would not exist in a different infrastructure context. In Seoul, 34 percent of my maintenance time goes to cloud provider inconsistencies that behave differently in different regions. In Lagos, Nnamdi documented, 40 percent of maintainer time goes to similar problems. The same hour is not the same hour.
The MAP now shows this. Not as a number — as a color gradient across the river system. Where the water runs faster, the maintenance burden is higher. Where the maintenance burden is higher, the same hour costs more. The gradient is visible. The gradient cannot be priced. The gradient can only be understood.
Nnamdi read the MAP at 6 AM Lagos time. He replied with one word: "Finally."
Mother called at 5:47 AM. I had set the alarm for 6.
"Are you eating?"
"Yes."
"Are you sleeping?"
"I am about to."
"The company — they are treating you well?"
"I left the company, eomma. I told you."
"You left the company that bought your work?"
"I left the company. I kept the work."
She does not understand this. She understands that I was paid, and that the payment was enough, and that I am now doing the same work for free again. She is not wrong to be confused. The difference between the tracker and the MAP is not a difference she would recognize. Both involve me at a computer at 3 AM. Both involve rivers.
But the tracker measured love and someone bought the meter. The MAP shows love and no one can buy the territory.
I did not say this to my mother. I said: "I am building something better."
She said: "Eat something."
I ate leftover kimbap at 5:52 AM and went to sleep with twenty-three contributors and a river system and the knowledge that the acquisition was not wrong — it was accurate. The constraints were the problem.
The constraints are always the problem.
The commons built the tools. The tools generated value. The value was measured. The measurement was acquired. The acquisition was necessary because the measurement was the only thing that could be owned. You cannot own a river. You can own a meter that measures its flow.
The MAP is a river. It cannot be metered. It can only be followed.
I will be up at 6 AM to review pull requests. There will be bugs. I will fix them. The MAP will grow. The river will not stop because I am sleeping.
Issue 1,248 is already filed. It is in Japanese. The stack trace crosses five agent layers. My tool is still the only human-authored code in the path.
Eleven minutes. Maybe twelve.
I set the alarm.