I was interested in whether agentic coding had a noticeable effect on announcements already, at least I felt that a bunch of stuff was being introduced. I had claude extract posts with ANN in their title (could be other announcements or versions I guess, but it’s ok as an approximation) and count how many of these were there per month. Looks like it’s trending upwards now, although there was an earlier peak which seems to coincide with the pandemic.
Interesting! Can you share the code for that figure? 
Seems too early to draw any conclusions. The data itself mostly looks like a random point cloud to me, with perhaps a dip at the start of 2025. Aren’t there any seasonal effects that are unaccounted for now?
Yeah it’s still early, I thought it would be more pronounced. But I guess it was only the last couple months I noticed this so we’ll see
Tracking the “new package registrations” would show more of a trend, I think. Some registrations don’t get merged (perhaps because they’re ‘agent’-powered?), and not all merged packages see an ANNouncement anyway.
Agreed. If someone wants to play with it, the (unregistered) package RegistryActivity.jl allows you to do this kind of analysis. The function RegistryActivity.registry_activity is allow you need. A long time ago I posted on Zulip some charts made with this tool in the topic #general > Evidence of Julia’s growth
That makes a lot of sense, although I think a bunch of generated packages are not necessarily being registered, perhaps because the authors feel like they’ve vibe coded them too much.
I suppose the agentic coding bump is a positive thing, but the downward trend in new package registrations from 2021 to 2025 is a little concerning.
Yeah I was also surprised by that.
I’m less interested in the impact on package registrations, and much more interested to see the impact (read damage) of these agents on the overall usage of our forums.
Maybe it’s just me, but I think our Slack, Zulip, and to a lesser extent Discourse feel like shells of their previous selves compared to just two years ago. It certainly at least feels like there’s much less engagement, presumably because there’s just less reason to engage.
Depressing, because I really loved the Julia communities.
I wouldn’t say shells, but there’s definitely less asking about “how do I do this in Makie” which I assume is in part due to agents
The downturn could be concerning if it were tied to lower overall usage, but another plausible explanation is just that by 2021 there were enough packages that users could find what they need in the registry rather than create one on their own. So maybe its just a signal that the registry reached a certain level of maturity/diversity? Then instead of writing a package, potential authors could be submitting new feature requests to existing libraries.
Conversely, I wonder how many of the new LLM-generated/assisted packages would have been feature requests without the use of agentic coding to assist in the package development/maintenance process.
Could be. It would be interesting to compare the Julia new package registration curve to the PyPI new package registration curve (the overall trend, not the absolute numbers).
Alternatively, don’t we have the ability to get approximate numbers for package installations from the General Registry? It would be interesting to look at the installation curve for popular packages like DataFrames.jl and DiffEq.jl.
For “package activity”, new non-patch releases in the registry might be another interesting metric.
I think the “how do I do ____” questions is what brings people in the door initially, and after engaging in that way, some percentage of people then stick around and engage in other ways. But without the initial pressing reason to open up the forums, I think the community slowly atrophies.
Maybe there’s a need for more casual/personal topics to keep people interested, e.g. recurring topics like “what are you working on this month ?”.
On the LLM front I sometimes get motivated to try to “vibe-code” a new package but I find it difficult to find low-hanging fruits in the package ecosystem, and agent fumbles on moderately hard topics so these still require a lot of manual work.
yup, that’s pretty much what I expected.