@ohdearlordylord: Instead of .md specs, try this: make the system a state machine (or rather, a reducer). Ask your slopus to make an inter…

The simulation is faithful, meaning it runs through the same logic as the production app. The specific diagram above describes the whole process of an agentic code edit to real-time semantic enhancement, classifier call and advice delivery right into the same agent turn.

What makes it more interesting (although I'd say the "flat" diagram already gave 90% of the benefits) is that when you model your code that way, it's easy to run automated simulations to check the system's "liveness" and also to show shared resource usage with concurrency.

LLMs can also reuse your exact production logic for one-shotting video games. This could be a useful form of educational material or a way to explain certain concepts in docs - in this case, queues and the most important nodes, e.g. Jev classification calls and delivery to agents

Pretty sure we'll see more and more of that in development: having better ways to visualise and understand LLM outputs is crucial to improving the understanding bottleneck. You can speak with your own Slopus about these approaches, analysing this codebase

Made with @bendlang https://bend-lang.com and @EffectTS_

Good idea

This looks neat. Helpful for you or the model?

It's about both: controlling model intent *and* discovering (and learning) the domain

Im a sucker for visual effects. Great idea

We stand on the shoulders of giants; the TigerBeetle approach influenced me here. I also believe we're in for a treat - I see more and more "I made my docs visualised like dwarf fortress"-type of posts. [maybe my bubble, though] People are going to interrogate agents visually.

How much tokens did you burn on this? Because i did something similar for work brainstorming. It landed on a similar reducer with select presets for different flows/behaviours but the diagram design skill can be quite restrictive in expressing the entire ER in detail.

One diagram won't express the whole system, of course, but it is still a great lens into its inner workings. That reminds me of the parable about an elephant and the blind sages. Token-wise, as soon as the model gets to see a verified event stream, I found Astra to be pretty good

Does it drive the real production logic?

Yep. That's the point!

What skills are required to generate this kind of reducer?

Mostly software engineering skills 😆. That said, I had some related agentic skills like domain driven design one installed, but that's probably it. I currently delegate to Astra anything that goes with spatial thinking (so, diagram layout and ux help).

bro is actually cooking fr.

Do you know TLA+?

bro this looks amazing im gonna have to try it out ❤️

Yes. And of course, you also have to add sound effects and upbeat background music.

Tried this - it’s a good way to diagnose but I felt it’s not a good way to understand/improve the architecture as slopus or the likes have a lot to heavy lift in terms of what these events mean what they could mean and what direction could the system take. Maybe skill issue at my

This is brilliant, friend!

有道理

not sure what slopus is but the idea is great for systems that become increasingly complex for humans to understand

状态机这个方向我挺买单的，尤其是模拟是忠实的这点，这等于把测试也顺手做了。但有个疑问，复杂业务里状态爆炸怎么办，靠拆子状态机还是一开始就控制粒度？粒度太粗的话，图又退化成另一版.md了

Rather than just reading the specifications of the article, it seems like you can catch omissions or assumptions in the state earlier by running the actual event and Reducer and turning it into a traceable diagram. Being able to ask "why" right there on the spot also makes it

State machines turn agent output into something you can inspect, test, and safely change.

cool 👍

awesome!

This better than html I use for repo docs

State machine reducer with interactive diagram beats static markdown specs

Works on tiny toy projects only

this is what I've been looking for, thank you. can you drop your prompt?

anyone can say anything these days

It’s called beads @grok

Some mistakes you might be making: Your effort level is too high Your skills are too long You're not aligning before you start work You're aligning too often, and too hard Your automated checks are too lax Your custom tooling is too token-inefficient You aren't doing enough

ban and delete most if not all code comments! agents love copying what already exists. a seemingly harmless comment explaining some temporary workaround is one of the worst viruses you can have in your codebase. like a weed it'll spread quickly as a tool for agents to quickly

it's crazy how easy it is to spot 10x engineers these days for a certain set of people LLMs totally unlocked their potential to the point where imagining their life before this seems totally depressing

New on our Engineering blog: Creating in-house source control. In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale.