OpenAI has begun measuring how coding agents are accelerating the work of AI research itself, examining where these systems excel and where their capabilities prove limited. The research, presented as 'Research acceleration: The jagged frontier,' captures a pivotal moment: machines increasingly capable of assisting in the design and development of more advanced machines. This work reflects OpenAI's direct interest in understanding the practical mechanics of recursive self-improvement—the feedback loop where AI development tools create conditions for even more capable AI systems.

The findings reveal notably uneven performance. Coding agents do not perform uniformly across research tasks. Some domains experience meaningful acceleration in research velocity while other tasks remain constrained by agent limitations. This performance variation—the jagged pattern that gives the research its name—proves instructive. Rather than a smooth exponential acceleration of AI development, the picture is more granular: pockets of genuine capability gains surrounded by clear boundaries and constraints. Understanding which research activities see acceleration matters for predicting how AI development itself might scale.

The implications stretch beyond research productivity. If coding agents can meaningfully accelerate certain aspects of AI research while simultaneously revealing performance ceilings, that shapes OpenAI's internal thinking about scaling and capability development. For the field broadly, this research provides concrete grounding: machine-assisted AI development is happening, it delivers uneven results, and isolating where and why remains essential. The gap between accelerated and constrained tasks reveals the real shape of the frontier.