Contribution Graphs Don't Matter Anymore

A dense green GitHub graph used to mean something. Not much, if we are honest, since it never showed quality, never showed private work, and could always be pumped up with trivial commits. But it correlated loosely with “this person ships,” and in a screening funnel starved for signal, loose correlation got used.

Agents ended that. Someone running coding agents daily can grow a contribution graph by something like 10x without becoming 10x more valuable. In my experience, that is roughly what daily agent use does to one. The level of the graph now mostly measures whether a process ran.

Which sets up a tempting move, and I know it is tempting because I made it. If the level is meaningless, read the derivative. Find the month the graph bends, and you have a timestamp of real AI adoption, checkable against the candidate’s story.

Claims eighteen months of agentic workflow, graph bent in March? Interesting conversation. It feels rigorous. It feels like signal recovered from noise.

It does not survive contact with the adversary. Graph-painting tools existed before agents did; there are web apps that will backfill a year of green for you at your preferred intensity. Faking an inflection takes a cron job and an afternoon. And Simon Willison has pointed out the deeper problem: an agent can now produce a convincing repository, commits, tests, polished README and all, in about half an hour. That is why he no longer trusts any of those surface signals and instead asks whether somebody has actually used the thing.

Everything that made the level fakeable makes the curve fakeable. And the moment advice like “look for the bend” circulates, candidates manufacture bends. The proxy dies on publication. This paragraph is participating in its death.

Everything that made the level fakeable makes the curve fakeable. The proxy dies on publication.

What the bend is still worth

One thing, and it is worth being precise about what. The bend is a question generator. “Your public activity jumped last spring. What changed in how you work?” You are not grading the graph; you are grading the answer, and answers have texture that cron jobs do not. Real adopters respond with specifics and, reliably, with complaints: the tool they started with, the loop they settled into, the first time an agent confidently broke something, what they stopped doing by hand and what they refuse to delegate.

Someone performing adoption reaches for press-release language instead. The graph’s remaining value is exactly this: it leads you to a question that is hard to answer from a script. The graph itself proves nothing in either direction.

In either direction, and the flat case matters more. A sparse or missing graph fits with heavy agentic work: private repos, squash-merge employers, contribution settings, company policy, GitLab, a self-hosted forge. Some of the strongest engineers you will ever screen have public graphs that look like a bad ski season. The rule worth writing into your process is blunt: no candidate is ever scored down for a flat graph. If your screeners cannot name the innocent explanations from memory, they should not be looking at graphs at all.

What replaces the proxy

Nothing convenient, which is the honest answer and the actual thesis of this piece. Willison lands in the right place: the signal that survives agent inflation is evidence that somebody operated a real thing. A tool they built and have used daily for months. A system they can walk you through at the level of decisions and regrets.

A plan document or agent session they can narrate: here is what I asked for, here is where it went sideways, here is how I caught it. Ask for provenance, in conversation. If you want a structure for that conversation, grade the loop against a ladder instead of reading squares on a calendar.

I keep noticing the pattern underneath, and it is bigger than GitHub. Every legible, at-a-glance signal in hiring is collapsing at once: the take-home, the contribution graph, the polished side project, the algorithmic screen. Agents inflate them all, cheaply, and each time we invent a cleverer proxy, it lasts exactly as long as it stays obscure.

The funnel keeps reaching for a dashboard because dashboards are cheap and conversations are expensive. But judgment about people, like judgment about code, turns out to be the part you cannot outsource. Read the inflection if it is there. Then close the tab and talk to the human.