The Lab

Repos


What a tracked panel of 66,704 developers actually put their hours into. Not stars. Pull requests and issues, counted monthly against 1,639 AI-tooling repositories, each labeled by a language model with the problem it's trying to solve.

Panel window JanuaryJuly · refreshed monthly · last collected Aug 18, 2026

1,639repos tracked
66,704accounts watched
31problems named
7months of history

The half-year in one picture

Share of classified panel attention, February against July. A line sloping up means the problem took a bigger cut of the crowd’s work, not that more people arrived.

FEBJULagent-harness48.2% 0.80×code-generation27.3% 0.93×multi-runtime26.1% 0.76×inference-cost22.4% 1.55×memory12.1% 1.80×tool-use11.8% 1.10×fleet-ops11.5% 3.22×supervision-burden11.5% 1.19×agent-native-access10.9% 1.48×context-engineering8.7% 0.89×observability7.8% 1.25×config-transplant7.4% 0.97×local-models7.3% 1.03×privacy6.7% 0.31×

What the half-year shows

The crowd stopped building harnesses and started running fleets. agent-harness is still the largest single problem in the set — 425 repositories, 48% of classified attention in July — but its share has fallen steadily. The work did not leave; it moved one layer up.

fleet-ops grew its share 3.2× between February and July, the largest move in the data, and memory grew 1.8×. Those two are the story of the half-year: once one agent works, the problem becomes many agents and what they remember between runs.

The clearest decline is privacy, down to 31% of its February share. Read that as attention rather than concern — the opening cohort carried a burst of that work which has not repeated.

These readings come from a rolling panel of 66,704 accounts, all polled across the full window; 20,330 recorded any activity inside it. How this is measured