Add page on agentic coding

[adrhill]

In theory, agentic coding could lead to a lot of new contributors to the Julia ecosystem. In practice, Julia's startup cost makes it awkward to use with agents, as they favor calling julia -e over using persistent background sessions.

To circumvent this issue, a bunch of tools have been developed, which can be broadly split into CLI and MCP tools:

• CLI:
  • https://github.com/tecosaur/DaemonicCabal.jl
  • https://github.com/dmolina/DaemonMode.jl
• MCP:
  • https://github.com/aplavin/julia-mcp
  • https://github.com/theogf/TestPickerMCPServer.jl
  • https://github.com/MichaelHatherly/REPLicant.jl
  • https://github.com/hexaeder/MCPRepl.jl

Which approach should be favored remains a topic of debate, and the field is moving quickly.

MoJuWo should give people pointers on how to utilize agentic coding, as most of this information is currently split over multiple discourse threads.

[adrhill]

If you use agentic coding with Julia, it would be great if you could leave a comment describing your current setup and what works / doesn't work for you!

[aplavin]

• MCP vs CLI: for most things, just a cli makes most sense, I agree. Persistent long-running session are one exception: as far as I know, mcp is the only way to have fully automatic lifecycle management, with all julia processes running only as long as the agent session exists. That's the whole point of julia-mcp: install once, fully automatic afterwards.
• What works: basically everything :) Package development, data analysis, plots, tests...
• Some issues:
  • Agents aren't great at following instructions, and no matter what you write in agents.md or in tool instructions, it will still do manually add packages to project.toml (sometimes messing uuids and spending time to figure it out) or do Pkg.activate(".../test") + include("runtests.jl") and fail.
  • Agents write working julia code without real issues, but the code is sometimes more like "fortran/python style in julia". They tend to insert manual type conversion that clutter the code, even if none are needed – eg Float64(... number ...) or ustrip(... number ...) or manually extracting x/y for plotting instead of using recipes. In general, they don't utilize dispatch well, and prefer for-loops to higher order functions.

[nhz2]

I wrap mcp commands in bubblewrap otherwise the agent will use Julia to do things on the internet I don't want it to be doing, or start deleting and editing files it shouldn't be messing with. Claude code already does this for its built-in bash tool https://code.claude.com/docs/en/sandboxing but this protection doesn't apply to the mcp tools.

[asinghvi17]

A bunch of julia packages need specific skills / prompts in order for agents to perform at their best. Makie, DocumenterVitepress and the SciML ecosystem come to mind as particular cases of this.

But in general my use of julia has been through CLI (after a single precompilation run, at least for me TTFX has not been a major issue). It's quite handy to put e.g. makie or web based UIs in a background task that the agent can read output from, so that it's also inspectable by a user. Recently I've been using julia-mcp for a couple things which it's been pretty good for - but for the interactive use cases the lack of "background-ability" is an issue.

[adrhill]

Ping @ChrisRackauckas @ChrisRackauckas-Claude and the #generative-ai gang @Sixzero, @Cvikli, @svilupp.

[xgdgsc]

I find the Cursor Agent / Debug mode is already useful for julia without any mcp/skill configurations. I just tell it the command to run the script inside the current project or the tmux session name to use and it edits files and debug just fine. Maybe it could be a good entry point for newcomers.