I always had this idea after reading so much about Lisp that I was designed for AI, but kinda forgot about it with all the craziness of the last years. When I read on the article the eval and the implications of agents self generating their code, it just clicked. Looking forward to experiment with this! kudos to the author!
I’m working on Kern, a small, git native, make and unix inspired llm workflow.
Modules are simply folders, and the tool just reads from stdio and outputs to stdout. Runs are stored in simple text files with all the info of inputs, outputs and other metadata.
I managed to run a council using OpenClaw twice a day of different models that would discuss my personal projects, and send me to telegram ideas or suggestions for the next steps. It was quite cool and got very nice insights, but was burning tokens like crazy, so I stopped it.
The self-extending part is wild! agent mutating its own tools mid-run makes trajectory fidelity even more make-or-break.
On git hooks: checkpointing state is easy, but if replay isn’t deterministic, bisect over checkpoints is unreliable... you can get different states on replay.
Tool evolution is a brutal test case: if the predicate is “does this tool still handle edge X?”, it needs to stay violated once flipped, or binary search happily lies about the origin tick.
Genuine question: when a self-built tool regresses, can you actually reconstruct the exact chain of reasoning/commitments that led to it? The artifact is simple to diff, the decision trail behind it is where it gets nasty.
When an autonomous agent fails at step 40, the bug was usually introduced at step 12. The hard part is finding it. Logs tell you what happened, but they don’t let you bisect a trajectory the way you’d bisect code.
I started thinking about what it would actually take to make that kind of debugging mechanical. It seems to require three things: immutable traces, pure reducers, and violation predicates that don’t flip back once they become true.
The interesting part: remove any one of those invariants, and there exists an execution where binary search over the trajectory cannot be guaranteed to return the correct onset tick. I tried to sketch a proof of that.
Once that substrate exists, though, you get something fun: fork, diff, and cherry-pick over agent reasoning. The same operations Git gave us over code but applied to trajectories.
Curious what breaks in the argument, especially the impossibility claim and whether the predicate regularity assumption is actually realistic.
vichoiglesias.com [email protected]