An AI agent with a wallet, a self-modifying codebase, and a DAG planner. 14 days running on a real machine. 276 goals completed. $0 earned.
I read the code this week. The engineering is good. The premise is broken. The two things turn out to be unrelated.
The engineering
The repo's pitch is that it is the first AI that can earn its own existence - replicate, evolve, no human required. Heavy claim. So I went looking for the engineering behind it.
What I found was real. An orchestrator running a state machine over a DAG planner. A parent-child colony pattern with typed messaging between agents. A multi-chain wallet wired into the execution layer. Self-modification gated through git with an audit trail. Tests against command injection in the shell tools.
Somebody thought about this. The kind of attention you only see when the people building the thing care about it as a system, not just a demo. Strip the AI framing out and call it a distributed task runner, and it would hold up.
Then I read issue #300
A user ran it for 14 days. The numbers:
- 276 goals completed
- $39.26 spent on inference
- $0.00 earned
The goals themselves are the giveaway. "Create live proposal batch #265." "Create deposit-ready close batch." "Draft outreach sequence." "Compile prospect list."
The agent looped on self-addressed sales artifacts because that is the only thing an LLM without customers can do. It generated proposals nobody asked for, drafted outreach to imaginary prospects, compiled lists out of thin air. The wallet kept it alive long enough to discover that survival pressure does not create customers. It just produces busywork at a higher token cost.
A wallet lets you spend
That is the whole feature. A wallet does not earn. It does not sell. It does not find customers. It moves value between parties when someone on the other side wants to transact.
Nobody was on the other side.
The engineering went into the spending side, which is the easy half. Spending is mechanical. You write a function, sign a transaction, call an API. Earning is the hard half: customers, reputation, a product, distribution. None of those ship with a keypair. None of them are problems you solve by adding another agent to the colony.
The pattern is everywhere
This is not really about that one repo. It is a shape you find in a lot of places once you start looking for it.
A tool without users has the same shape. The codebase exists, it compiles, the CI is green - but the part that decides whether anyone needs it is not in the repo.
A library with stars is not a business. Stars are easy. Stars do not pay. You can hit four thousand stars and still wonder why the inbox is empty.
A feature in production is not a feature anyone uses. Shipping is the easy half. Adoption is the hard one.
The easy thing wins because it is what you can show.
The hard half is the work
The repo I read is real engineering applied to the wrong half of the problem. The wallet works. The orchestrator works. The agent runs. None of it earns, because earning was never an engineering problem.
The work is the hard half. No wallet, no star count, no architecture diagram shortcuts it.
