How AI Agents Built This, and What a Human Decided

This repository was built mostly by AI agents under one person's direction, and that is stated up front as the point rather than buried as a confession. Across 3,224 commits from 2026-03-09 to 2026-07-22, about 95.75 percent carry the agent identity and 67.9 percent carry a Claude co-author trailer; the busiest single day logged 275 commits, and roughly 4 percent of commits were hand-authored by the human. A reviewer does not have to take that on faith — it is recountable from git log on a fresh clone.

So what did the human actually do? The load-bearing contributions are exactly the ones a fleet cannot supply for itself: the validation methodology, the ship-or-reject decisions, and the honesty rails that make an agent's output trustworthy. The keystrokes are mostly the fleet's; the judgment is the director's. The high agent share is the feature, because it is the proof the workflow scaled without the human becoming the bottleneck.

The machinery is what keeps an autonomous fleet honest overnight. A fail-closed ship gate decides whether any candidate signal ships, and it is built to refute, not confirm. A signal is never tested in isolation; it must clear five criteria jointly — an expanding walk-forward where every fold must improve, a null-shuffle permutation control the real effect must beat by at least three standard deviations, an ablation that measures the signal's marginal lift when added to the full production feature set rather than alone, a calibration and coverage check, and a closing-line-value check. A Benjamini-Hochberg correction runs across every signal ever tested to control false discovery, and a final held-out set is touched exactly once. If no leak-safe data bundle can be built, the gate returns DEFER — never a false SHIP. That design is why the reject pile is so large: the gate was tuned to kill candidates, and it does.

Candidate ideas come from an LLM-free proposer that enumerates feature transforms from residuals and screens them with a cheap statistical filter, so discovery never runs out of hypotheses — but the gate still decides, so volume never becomes noise.

Two guardrails are enforced in committed code, not politeness. A pre-edit check marks the production trees (the engine, the API) as protected, so an agent that tries to edit them is denied and routed to human review. An honesty linter plus the no-edge rail ban dollar, ROI, and edge language from any output, so an agent physically cannot ship a profit claim. These are the mechanical reason the fleet stays inside its lane while the human sleeps.

The system is also honest about its own size. It is roughly 430 Python modules, of which about 130 under the prediction folder are a research surface with only around a dozen load-bearing in the live deployment graph; the API layer serves about 99 endpoints across 12 routers. Being explicit that most of the surface is experiment, not production, is a judgment claim, not a size brag.

The current-era senior signal is not can you write code — a fleet writes the code. It is whether you can direct that fleet so the output is trustworthy: route work by cost, run executors in parallel without corrupting the repo, and gate every result behind validation strict enough to reject your own favorite idea. The human wrote the gate, the guardrails, and the methodology; the agents wrote the implementation; and the split is recorded in git, not asserted.

Sources