Julia core sees agent-written PRs surge, community frets

At this year’s State of Julia talk at JuliaCon, core developer Keno Fischer said agents had authored roughly 80% of pull requests to the main Julia language repository in recent weeks, that he had personally burned through about $45,000 worth of Anthropic tokens in three months, and that contributor workflows should treat an agent pass much like running tests before a human submit. The numbers landed hard enough that Ronny Bergmann and others opened a long Discourse thread on whether heavy agent use is compatible with a healthy volunteer community.

Not everyone can or will follow that path. Frontier models cost money many contributors lack; access can be cut off by vendors or geography; some simply prefer writing Julia by hand; others object on ethical or environmental grounds. That raises a practical question the thread keeps returning to: how the project treats people who stay agentless if agent review becomes the expected bar, and whether pooled or subsidized access can close the gap without baking in dependence on free compute that may vanish.

Supporters point to speed and quality controls already in use elsewhere in the ecosystem. Chris Rackauckas’s JuliaCon talk on SciML’s agentic maintenance bots offered quantified evidence that agent-heavy work need not mean “slop” when paired with strict inference and testing gates, though critics noted such setups can rest on hundreds of thousands of dollars a year in complimentary industry compute. Others argued agent review is an optional extra check: if the service disappears, the project loses a filter, not a foundation, and can adopt it later once results on real PR volume are clearer.

Skeptics worry about softer damage. Help channels and forums already feel quieter; if newcomers learn by prompting rather than asking humans, recruitment and mentorship thin out. In fields such as bioinformatics, some see a drift toward “Rust for everything” with agents writing the hard parts, a shift that favors languages built for static guarantees over Julia’s interactive strengths. Defenders counter that multiple dispatch and trusted shared packages still beat one-off vibecoded stacks, and that faster core development may clear long-standing blockers such as latency and deployability.

Moderators have asked participants to keep the tone civil and to time-box the thread to about a week rather than let it run indefinitely. The underlying tension remains unsettled: how far Julia should lean into agentic throughput without pricing out contributors, hollowing out the social layer that sustains an open-source language, or betting the house on subsidized models whose economics are still unstable.