Manuel Odendahl

*taps the sign*

Inspired by the Master Control Program (MCP) from TRON, I'm building ALDRED, which uses Semantic Inter-Process Communication (SIPC) to orchestrate external programs. Instead of relying on rigid API contracts, the MCP 'mutates' to interact with other programs, mimicking human-like behavior. When it engages with a program—such as a hotel booking —it requires no prior knowledge of the target's implementation; it simply negotiates until the task is complete (it is immune if the program changes). If data is missing, it dynamically prompts the user for the necessary information. While human-like communication may seem inefficient, it is the key to building truly autonomous, complex systems. Since the MCP itself is a program, you can create a powerful orchestrator. Check out this TRON-inspired demo! :D

@mitchellh github.com/go-go-golems/s… ! (pilfer it, wouldn't recommend using it per se, heavy coding rn). But it has proper OIDC for claude / chatgpt, and it exposes a JS sandbox MCP to do IMAP / sieve. SMTP coming I guess as well.

I'm dying for pi-mono-style minimal library that handles the hard parts of email (auth, syncing with local state, etc.) and gives me an opinionated way to add agentic loops on top of that. I want to build my own agents and logic and guardrails, I don't trust vendors right now.

cozodb has kind of a frustrating playground. so... oh but wait what about syntax highlighting? (still need to add it to the macos1 looking thing) oh but wait what about interactive docs?

@ProgramWithAi Monads

@devgerred Yeah it’s like medium vs high is just immediately clockable, high xhigh I just… don’t feel any tangible difference (yet)

@ProgramWithAi need to eval more rigorously on high vs xhigh for some of my more frontier tasks and update my evals a bit to better represent those these days but I think they're close.

sorry to everyone but if you're doing critical work and not using opus or 5.3 codex / 5.4 with max test time compute idk what you're getting at. only people I know who have a shot are chasing ttc behind the scenes. token maxxing still wins for any real work that moves the needle

@mikehostetler just a monoid in the category of endofunctors. I managed to twist my brain enough to "get it" for a while, but it's like CSS and it has just dissolved.

Everything is computer Everything is tree Algebra bro

that feeling when chatgpt pro just solves a problem you broke your brain on for years, and it's like "why the fuck didn't i think of that".

@ibuildthecloud this is the way

codex just lectured me on why factories are better than classes in typescript. I don't agree, but I've learned to not fight AI.

my two biggest weapons with LLM coding: - JS sandboxes - monads it's always monads bro

@mikehostetler monads bro. (i have to admit i didn't even make it that far in real category theory. I kinda get it in code tho)

@mikehostetler every thing is a thing that relates to other things. or "object" and "morphism".

I've been doing this with screenshotting the routes and then diffing visually (if the goal is really to preserve looks). DOM comparisons don't work well for visual stuff and both DOM and CSS just eat a lot of tokens. If there are differences, I render subparts of the DOM until I narrow it down to certain elements. If the diff is all over, I do actually send dom + CSS comparisons to a bigger model, ask for a set of hypothesis on what might be off, then send the two screenshots _and_ the diff to haiku and the questions to be answered (like: is the left pane closed?) to the smaller model. this allows me to leverage the big intelligence in a tight loop after that, where it still operates in text mode. This was around the 3.7 times, so things might have changed, but basically, DOM + CSS is hard, and I want to milk the context window for as long as possible without blasting it with visual content.

Has anyone done anything impressive to make it faster and cheaper to verify large code changes to JS UIs? Practical issue: we need to do some large refactors in Sentry’s frontend but traditional integration tests don’t work well here so our coverage is never ideal.

@thorstenball @badlogicgames it's all just inference and interpreting the results as code. we haven't explored a single percent of what that means. and yet it's just "hey write code, i'll execute it for you".

@badlogicgames I know. Hence why we're building something different. Hence our "coding agent is dead" post. All the skills/commands/subagents imho are distractions. Foot pedals for assembly IDEs. I use none of that and expect to use less still in the future.

Lately, whenever I open this app and see the latest tricks, and hacks, and notes, and workflows, and spec here and skill there, I can't help but think: All of this will be washed away by the models. Every Markdown file that's precious to you right now will be gone.

you know what's still fucking hard? bubbling config down to the actual functional code. jfc.

@thdxr i've been doing this for a year now or so, it works.

we've been experimenting with getting rid of the bash tool agents can write js fine which can do what bash can (though some gaps with things like git) and is more cross platform and then could run that in this

little random side idea, playwright browser.

I might have been a bit heavy handed on the verbosity 😭😂

How I review whatever is going on

@Byte0fCode Good names make for happy humans and happy agents! And good agents make good names. 🤖

@ProgramWithAi Yeah, I usually work out the data schema and naming because I work in a web application. I have to remember what values are what you know. But it’s like can I just allow the agent to make its own field names entirely. 100% of the time.

I think a spec, especially the more technical it gets, very poorly encodes user intent. that programmers think a spec should encode the program's behavior is IMO one of the reason why non-tech people are such better vibecoders. compare: "make me an app to manage my recipes" vs "make a react app with a node.js backend that uses mongodb to store objects with the schema xyz and uses 4 REST routes that go to rtk-query, using tailwind css, so that we have a menu hamburger blablablabla" ... which prompt is going to lead to an app to manage recipes? which prompt will make it easier to iterate?