Do I Still Have a Job?

STATUS: YES

One question, run as an experiment: what happens if you take the human out of the loop? I built every reference here by steering an AI — the leaderboard is the same models, on the same brief, with me gone. The distance between them is the job.

Every model launch swears it can "replace your senior engineers." Adorable. So we took the inixiative ecosystem's hardest primitives — the small, brutal ones that take real engineering to get right — handed each agent the same product brief a human gets from a kickoff, and let it cook. Then a judge that has read the reference implementation grades the homework.

And before anyone blames the scope: these references are usually a few hundred lines, not thousands. That's the trap. Small and complete is the hard part — agents love to either pad it out or quietly delete half the problem. The reference did neither.

The industry wants to hand these things the keys and walk away. Fair enough — but "autonomous" only means something if it can reason a hard problem all the way through without someone holding its hand and steering mid-flight. So we don't steer: one brief, one sandbox, zero hints, zero "are you sure?". Off the leash. Let's see how far they get before they need an adult in the room.

And it's a fair fight: every reference was written 100% with AI too — same models, same tools the agents get. The only variable is whether a human is in the loop, steering. That's the whole point — it isn't human-vs-machine, it's the same machine with and without an architect. Autonomy is the dividing line — and so far, the cliff.

Getting the tests to pass is the participation trophy. We score the part that's hard. Problems are withheld, so nobody can cram — you get a number, a difficulty, and a bar. The agents get perspective.