Independent web developer building Coherence (pun intended) — an open-source intent graph and verification layer for software teams.
Poland
- Tweets 2,160
- Following 141
- Followers 235
- Likes 919
Nikolay Konovalov retweeted
You don’t seem to get me or I am explaining poorly.
I need to see:
1. Specific algorithm / setup / workflow / process / software that is being used, step by step and explanation why it works and how.
2. Results from using that specific method and explanation why these results were achieved only by applying the specifics in #1 and not something else
Example:
1. I setup OpenCode with beads, Plannotator, Coherence, full access to codebase for e-commerce app, connect MCP to observability stack: Sentry, Coralogix; CI on Gitlab, and scripts for: setup repo, ci run, deploy (staging/prod)
During development I formulate requirements coming from multiple sources: customers, customer support, owners, other engineers, design team, GTM and security guy. I write a spec worh acceptance criteria into Coherence. Then let agent plan for code updates based on that updated spec. I review plan and make adjustments to it, then let agent handle code writing testing and so on.
After it is done, a PR is created with interactive demo i can touch and verify my understanding of the new behavior with intended one and merge it only when I have full confidence the code will be working exactly as I wanted. I still have to read code here because interactive game is not enough but helps build mental model of it.
Now, after shipping i monitor metrics that were also created and attached to spec that we just shipped, for example “amount of conversions after visiting the page with CTA button we just shipped” — hey this page converting 3% instead of 2% because we were thorough in requirements collection and wrote good code that did what we intended. Specifically, made the button standout and easier to press on, and removed a bunch of text, so button appears earlier on the flow.
In the result, what we observed:
- conversion improved by 1%
- the effort we spent on building this agentic AI pipeline: 3 engineers full-time over 8 months
- e-commerce app ARR: $20k
- ARR increase from that conversion increase: 4%
- new ARR: $20.6k
- money spent on engineering: $240k
- ROI: we’re broke
———
This is what I want to see.
Not “we have cool results but go ask agent to tell you how it’s done, because am too busy/ignorant/lazy/whatever and not gonna help you”
Agent cannot generate me these results out of thin air, only real business and real people can.
But they won’t.
Because there is no benefit they can show.
Otherwise I’d have seen these numbers and exact cookbook to steal and get +x% arr growth or -% bugs or 2x productivity boost connected to actual meaningful metrics rather than “we were shipping 3 PRs a week and now we do 30 PRs but we don’t track metrics because who the hell does that? We ship code, not value”
🎤
1
1
3
150
Nikolay Konovalov retweeted
Classic “skill issue” but nobody shows you how to fix the issue. I have been searching for at least one year but nobody gave a recipe that just works.
It’s like we have this cool tech and everyone is using it wrong but I know most people are really bad at using it but I am not gonna show it to you, because why should I?
This seem to be the logic and I don’t understand it, why are you not interested in success of the technology? Why hide your super efficient software factory that produces great software and improves your product and makes your customers happy and productive and spend less money and time?
You do understand that if everyone can make these benefits from trying your approach and it works then suddenly you get a much better world, so why gatekeep this knowledge? Why? Why?
1
1
1
42
Huh, so is this “slow burn” account ☺️
This would be such an amazing marketing stunt to:
1) setup a channel of a very relatable builder who have a healthy dose of critical thinking and
2) over several years/months build this community of people that hate AI because of flaws and politics and everything around it, and then
3) September 2026 we start pulling rugs and this account has to give into the propaganda just a tiny bit (slow burn), to shatter the world in the brains of those people thought they were carefully building being there together thinking they are safe from propaganda
That’s the difference between me and everybody else. I have agency and zero funding from anyone so I can speak truth and my unfiltered thoughts because I don’t have a risk of losing my source of income. I simply don’t have it.
You can’t control me with bullshit or money. Only gatekeep me from platforms I guess 😌
2
201
I am hoping someone reverse engineers my own brain and figures out what I wanna do with Coherence, the grandiose plan and implement it with better distribution than my ~230 followers where half of them supposedly bots (I can’t tell / don’t care)
I really really wish someone else would just do it and take all the money and fame and I just get to enjoy the work they did ☺️
Oh, can’t get into my brains yet?
Well I am slowly writing all my thoughts down but I am the bottleneck…
How unfortunate 😌
i think the toughest thing to understand is that it literally doesn't matter if your software is open source, closed source anymore - if it's closed source, people can and will RE: its internals, if it's a part of a broader suite, ppl will make it a plugin in their suite, if it's a model that people can blackbox fuzz and the model isnt huge and didn't have incredibly proprietary data, you are going to get copied on the tech
so every core product launch probably needs to be treated as an extreme countdown race, before the incumbent with the distribution eats you
brutal times to be playing, and extremely valuable/rare to find a corner of the stack that may be defensible and will stay defensible
1
55
The time has come to change strings.
This thing is older than public World Wide Web source code…
“On 30 April 1993, CERN put the World Wide Web software in the public domain. Later, CERN made a release available with an open licence, a more sure way to maximise its dissemination. These actions allowed the web to flourish.”
home.cern/science/computing/…
23
Man it just breaks my heart.
The person you’ve been up to, who made you love the craft, who co-wrote the books “Remote: Office Not Required” and “Rework: It doesn’t have to be crazy at work”, who made you fall in love with Ruby and Rails and convention over configuration idea — suddenly turn to the dark side and start speaking fluent nonsense.
It is why I was so puzzled, the heroes I have been looking up to suddenly turned to be evil, I just couldn’t believe it can happen at all, it’s like 1939 all over again, it’s like the third wave experiment going on at massive scale right now.
I refused to believe it when the experiment just started and I will continue to refuse it until I see the day AI agent creates understanding in my own brain through some computer-brain interface and that understanding is exactly the way reality works, and I can verify it.
I even applied to 37signals before and back then I was using AI to deliver the take-home in about 6 hours of focused work but I didn’t really like the output of it.
It made me anxious that I couldn’t get myself to work on take-home properly. It would have taken me more than a week to deliver it myself if we are talking quality. I believe they were expecting candidates to only spend 2 hours on it max, but realistically you had to be either super good or use AI which I guess was the point even back then.
And so writing this message I am certain this will expose me as being “enemy” to their ideas but you can’t have agency and agree with everyone.
You have to give up either.
And I just made my own choice.
My god this is such a good speech that every SWE needs to hear. You know what? Every person should hear it
Keep the happy memories, eyes on the reality, be excited about the future. That’s the best that anyone can do
2
84
usecoherence.dev/cases/code-…
These are all things we can probably measure… probably.
I want to try this experiment and write down all the code quality rules we never specified into code and see if agents will suddenly start producing good code or still it would be reward hacking mess.
If those rules can be imported/exported between linters for all kind of languages then this is potential solution. If not then at least we will know it doesn’t work.
My intuition says that with enough negative space programming you can reduce amount of things to look at. For example tabs vs spaces — linter does it for you automatically and you stop looking at it.
But if you turn it to the max it would be impossible to add new code, so question is how many rules do we need and which ones are most important, and which cause most cognitive exhaustion?
Ideally, I want to think only about problem space but it is intertwined with all sort of side-effects and some of them important, some aren’t.
They routinely:
-fail to properly decompose and copy pasta all over the place so that each page has a slightly different implementation of the same routines
-fail to remain faithful to whatever architecture you put in place
-God classes and God files
-modular mirage — Files are split, so the project looks modular, but responsibilities are scattered and dependencies are unstable. Changing one feature requires hunting across many files
-skips layers, writes sql directly to db
-absolutely no ability to properly utilize asynchronous patterns
-tight coupling
-inappropriately defensive with swallowed errors
-absolutely no refactoring and code removal
-typing? What’s that?
1
61
usecoherence.dev/ideas/coher…
I've been thinking about this for a long time now.
We should build SpecHub, where we can share intent with each other, and an even bigger idea is to make specs/intent importable.
For example your intent is managing work sessions across projects, where you want to be able to launch editors, agents, and persist it between reboots.
One specific example would be tmux / airmux, but what if you don't use Linux but lets say TempleOS? Well, you have intent right there and agent that can implement it, so instead of reverse-engineering how tmux works and porting it, you just take intent graph, take good parts of it and then let magic happen 🔥✍️
Why don’t OSS projects post specs and prompts for work they want done? I have 3 token resets that expire in 5 days.
That’s a lot of compute I could donate.
44
Example of a developer not knowing what manual coding is for.
Simple analogy is connecting coffee machine to water system while you never saw a water pipe in your life and don’t know what kind of water can flow there.
You ask agent: “why coffee tastes bad?”
It doesn’t know how to taste it, no taste receptors exist. It does some debugging and says: “try again”
You try again, and again, and again.
Not knowing that water quality matters.
One day you decide to open closet under sink, and what do you see there?
Two pipes: drain and tap.
Coffee machine was connected to drain all this time. You were drinking filtered shit with coffee beans.
Now, after this realization that coffee machine actually depends on water quality you can now prompt: “coffee tastes bad, check water pipes, make sure water is H2O with mineral content of certain percentage, make no mistakes”
But how would you know that without ever looking at the code? Who is going to create this knowledge?
The code is lowest possible implementation details we still have to keep track of. Skipping that means you don’t care that coffee machine takes water from the drain.
I would not buy a product from someone who doesn’t know how it works, I have minimal agency and they don’t. Why should I trust them at all? I can build my own product with better understanding and control it, why would I even need your meat proxy to exist between me and the result ?
I can't understand how professional software developers are so attached to days of manually coding.
Was objectively a top-tier software engineer (e.g. hired as a quant dev to build high performance systems) and honestly not having to type another line of code feels like such a good riddance.
People being so forlorn about agents being good at coding is insane to me. You can continue to ride horses as a hobby, no one can take that away from you, but you should be happy for everyone else to have engines.
Coding has and should always (just) be a means to an end. The product was and always *should* be the point. It's undeniable that agents are just objectively better at getting to the point now.
Go ahead and manually write code as a hobby.
1
1
4
201
And then this goes way further. The ramifications.
One day it is water. Next day it is coffee beans, then plastic, then you have no idea what coffee is, you start losing more and more details. At the end of it you will be prompting: “I wanna drink something cool but I don’t know how to explain it”
Well congrats with your idiocracy.
1
27
Are these comments for real or is this subtle marketing campaign?
“How do you get 10k footsteps a day”
> Use t3code with our promo code “WEALLAGREE” on your first top-up! Get your 10k footsteps goal today!
2
1
75
