Agentic coding workflows

[dannytaurus]

@David-Healey It will probably be useful for refactoring and moving things into namespaces. I'm on a short HISE-atus (see what I did there) at the moment but I'm gagging to get back into it now we have the MCP stuff set up.

The future is leaning back in my chair with a cup of tea casually throwing out commands like "now add an effects section on the right with a Juno-style chorus and an authentic BBD-style delay, and wire them both up to XY pads with, I don't know, you choose the parameters."

[David Healey]

@dannytaurus said in Agentic coding workflows:

terminal-based approaches like Claude Code can't do that. When you finish the task and close the terminal tab/window, all the history is gone.

Does it not save it for you to view in the web version?

@dannytaurus said in Agentic coding workflows:

I also pay $20/mo to OpenAI for ChatGPT but now I'm trying out the Codex coding environment that all works within the same $20/mo subscription.

How do the Codex models compare to Sonnet (that's the main one I'm using)? And does it have usage limits like Claude (resets every 5 hours/week)?

@dannytaurus said in Agentic coding workflows:

moving things into namespaces.

Does not compute... you mean you don't start by outlining your namespaces before writing your script

[griffinboy]

@dannytaurus

Haha, we aren't there yet.
AI is a fool when it comes to Analog Modelling.

But maybe someday.

[dannytaurus]

@David-Healey said in Agentic coding workflows:

Does it not save it for you to view in the web version?

Oh, I don't know. I've only used the terminal version. Makes sense that there would be a web backup off chats though. I don't really get how it works, coding with AI on the web. I guess you just link it to a GitHub repo and then view the diffs in the web app? Feels a bit detached to me. But I am slowly spending less and less time in the actual files, and more time reading diffs.

How do the Codex models compare to Sonnet (that's the main one I'm using)? And does it have usage limits like Claude (resets every 5 hours/week)?

I found GPT-5.4 to be on par with using Opus 4.6 in Cursor. Mind you, I often switch to Sonnet in Cursor because I don't feel like there's a huge difference for what I do. And Opus costs 5X Sonnet, so that's good.

Does not compute... you mean you don't start by outlining your namespaces before writing your script

Not at all. I still haven't got my head around when and why to use namespaces. Maybe I haven't made a plugin complicated enough to really benefit from them. But that's what's cool about AI. I can code up ad-hoc from scratch then have the model tidy it up for me. Of course, that would be much more reliable if we had unit tests for everything

[David Healey]

@dannytaurus said in Agentic coding workflows:

I guess you just link it to a GitHub repo and then view the diffs in the web app?

Yep that's it

@dannytaurus said in Agentic coding workflows:

I found GPT-5.4 to be on par with using Opus 4.6 in Cursor. Mind you, I often switch to Sonnet in Cursor because I don't feel like there's a huge difference for what I do. And Opus costs 5X Sonnet, so that's good.

How does the price compare? I'm mostly using Sonnet for coding but I also have a ChatGPT sub, if I could drop the Anthropic sub that would be good.

@dannytaurus said in Agentic coding workflows:

when and why to use namespaces.

It's like a class in an OOP language. I group related code together. For example if I have a settings page on my GUI then I will create a UserSettings namespace. For the preset handling I will have a presets namespace. For laf I have a look and feel namespace, etc.

[David Healey]

Apparently Claude Code doesn't read AGENTS.md, you need to add a CLAUDE.md too.

Claude Code's /simplify skill is useful for cleanup after making changes: https://claudefa.st/blog/guide/mechanics/simplify-batch-commands

[Bart]

This post is deleted!

[JulesV]

Please correct me sorry if I'm wrong. But as far as I remember, when people here were talking about creating AI-powered plugins, they were referring to plugins that use AI for DSP operations, analysis, etc.

But how and why did we get to the point of writing code using AI? I find that weird.

Isn't writing code and creating plugins with AI a bad idea?

[David Healey]

@JulesV said in Agentic coding workflows:

Isn't writing code and creating plugins with AI a bad idea?

Vibe coding is a bad idea, but using AI as a tool to write code for you (which you then check properly) is a very good idea. It saves a ton of time.

[JulesV]

@David-Healey I understand, but isn't this technically vibe coding? @Christoph-Hart

Also, I think it seriously jeopardizes the future of the profession. Saying that AI can't produce good plugins while simultaneously training it to do so is, in my opinion, wrong and will pose a long-term health problem for those who make a living doing this job :)

I think AI could be used this wrongly :)) Wouldn't it be much better to focus on analytics and DSP topics, as I mentioned above?

[Christoph Hart]

But how and why did we get to the point of writing code using AI

Simple: the latest iterations of the SOTA models make AI coding on par with a reasonably well skilled developer.

[JulesV]

@Christoph-Hart said in Agentic coding workflows:

But how and why did we get to the point of writing code using AI

Simple: the latest iterations of the SOTA models make AI coding on par with a reasonably well skilled developer.

It is the begining of the end then :)

For those who think AI can't create good plugins, train your AI yourselves, and eventually this prophecy will finally come true :)

[David Healey]

@JulesV said in Agentic coding workflows:

isn't this technically vibe coding?

Vibe coding is where you prompt the AI and accept the output. When it doesn't work you give it the error message and accept the new output. You don't look at the code, you don't understand the code, and you don't care about its quality.

What we're doing it prompting the AI with a problem. We're examining and understanding its output. We ask for considered changes, or we make the changes manually (although I find I'm doing less of that). Then we test thoroughly and go back to the AI if it needs further changes. The code quality is maintained.

@JulesV said in Agentic coding workflows:

Saying that AI can't produce good plugins while simultaneously training it to do so is, in my opinion, wrong

AI can already produce good plugins, that ship has sailed. But they are not necessarily interesting plugins or widely useful. They are more specific and tailor made for individuals. Although there is no reason why you couldn't use AI to generate a saleable product - I'm sure lots of people already are.

But we are not using AI to create an entire HISE project (yet). We're using it as a tool to help our development process. It's like if you hire a developer and you tell them what you want, you don't write the code yourself, you just steer the developer in the direction you want to take the project and they do the actual grunt work.

@JulesV said in Agentic coding workflows:

Wouldn't it be much better to focus on analytics and DSP topics, as I mentioned above?

DSP and Look and Feel is probably what I will mostly use it for within my HISE projects. But I'm using it a lot for adding features I want in HISE itself, or fixing bugs. I'm sure Christoph is going to be overwhelmed with all my pull requests :) We'll need to train a Christoph AI to manage them.

[Christoph Hart]

Nobody is expecting that AI will produce interesting plugins on its own - the same as it will never produce interesting music with any artistic value ever.

It‘s the middle layer that handles the transformation from a human having an idea and vision to a plugin that is completely being revolutionized at the moment - this is where HISE sits and ignoring this paradigm shift would be suicidal for a framework like HISE which is why I‘m pulling 14 hour days since about 3-4 weeks to figure out the best way going forward. The way Opus 4.6 interacts with the current toolset I‘m developing is uncanny though and goes far beyomd simple LAF stuff - it eg. oneshotted the transformation of a fairly complex user preset model to the custom automation handler then Ralphed it‘s way to find out that setting processorIds will make the control callback not fire - with proper guidance and planning, which ties back to daves distinction between vibecoding and AI assisted development.