burnerToBetOut

Because the autonomous agent deal is new to me, this thread's blog post piqued my interest. I thought about trying to replicate the experiment. But I'm lazy :)

Instead, I ended up getting a coding agent to run through a thought experiment [1] based on the method described in the blog post.

The AI's critique sounds kinda harsh to me. So, I'm quoting only a snippet of it here…

———

…Running this specific recursive loop without a higher-level architectural constraint or a "convergence metric" will likely result in a specific type of technical debt known as _Ravioli Code_ (the inverse of Spaghetti Code)…

…

The prompt provided optimizes for _local code metrics_ (file length, shallow SRP) at the expense of _global architectural cohesion_.

In your specific $A \to B(C)$ scenario: _$A$ stops being a class that "does something" and becomes a class that "configures things that do things_."

The eventual system is one where every piece is perfectly unit-testable, perfectly mockable, and adhering to strict SRP, yet the system as a whole is incomprehensible to a human reader because the "story" of the code has been shredded into a thousand paragraphs scattered on the floor…

———

Again, those are the words of the automaton; not mine. I simply pointed it at the blog post and this thread. I asked it what it thought of the approach and final results discussed in the blog post.

Then, in the vibe coders' vernacular, I "let it rip" ;)

Today I noticed a follow up blog post [2]. That one noted a number of things that resonate with my relatively lightweight experience with agentic coding…

    > …
    > We're not there yet…
    > …
    > If you want the agent to tackle a
    > specific architectural smell, you
    > need to name it…
    > …
    > You need to give direction. The
    > refactoring principles are specific
    > to each project's goals and technology choices.
    > …

[1] https://g2ww.short.gy/TIL2Ralph

[2] https://frederic.vanderessen.com/posts/unsupervised-refactor...

    > …How do you keep architectural
    > intent explicit…
    > …
    > …without slowly drifting into a
    > system you no longer fully understand…?

One way some of the more mature shops I've worked at have solved the understandability problem (for human devs) is with visual modeling.

Visual modeling doesn't necessarily mean by-the-book UML, of course. Even junior devs understand simple misshapen boxes, lines and arrows on a white board.

You'd be amazed at how effective even imperfectly-drawn visual models are at keeping meatbags' memories on track with how the system hangs together.

In a brick and mortar dev shop, ideally you want a large whiteboard that's positioned somewhere that every team member can glance at conveniently at any time.

I know it works wonders for me, anyway. In fact, just yesterday I added a PlantUML-backed visual model generating skill to the coding CLI I've been using lately. They get output to a uml folder at the root of my projects.

    > …I’m curious how others handle this in practice…

I'm banking on good old fashioned UML defined as machine-readable text files as one way to remind coding agents of architectural intent in my AI coding-assisted projects.

Please clarify the Google connection.

I'm guessing that it's an official Google-built product. [1]

[1] http://support.google.com/notebooklm/answer/16179536?sjid=62...

    > …You don't need to fully explain the
    > concept as it probably nows them well…

Unsurprisingly, many would disagree [1]…

> 1 Establish a Clear Vision

…

You have experienced the world, and you want to work together with a system that has no experience in this world you live in. Every decision in your project that you don’t take and document will be taken for you by the AI…

[1] https://news.ycombinator.com/item?id=46916586

    > …our services interact with http apis…
    > … 
    > …If it was internal apis or libraries…

That reminds me that I wanted to ask you: How good is your agent with complying with your system's architectural patterns?

Given my admittedly limited experience with coding agents, I'd expect a fully autonomous agent to have a tendency to do naïve juniory dev stuff.

Like, for example, write code that makes direct calls to your data access layer (i.e., the repository) from your controllers.

Or bypass the façade layer in favor of direct calls from your business services to external services.

FWIW: Those are Java/Spring Boot idioms. I'd have to research whether or not there are parallels in microservices implemented in Go.

Implicit knowledge like what you know about Opus 4.6 (and that I don't) is what I meant about it being "an amazing learning opportunity".

So, thanks :)

    > …If you don't measure it, you don't see 
    > it until something breaks in production…
    > …
    > …the slow, silent divergence between
    > what you intended and what actually exists…

What's your take on the absence of any mention of tests in the OP's loop steps?

Awesome! Thanks again!

    > …Read the TASKS.md file if it exists…

What about sharing that TASKS.md? I'd like to replicate your success as closely as possible.

Having the same tasks would help me nail a similar successful result.

I forgot to also ask: What language and which REST framework was your microservice implemented with?

For even better reproducibility, I'm thinking I should have my codebase be as similar to yours as I can get it.

TIA!

    > …to refactor and improve the quality of a microservice.
    > …
    > It worked. It vastly improved the code base…

That sounds like an amazing learning opportunity!

Would you be up for sharing before refactor and after refactor versions of your experiment in a public repo?

Thanks in advance.

Please do an Ask HN: Who wants to be fired? (February 2026) next! ;)

Being serious though. Seeing 300+ people (and counting) effectively doxing themselves so eagerly is something I never expected I'd witness in such a tech savvy community. [1]

Desperate times ¯\_ (ツ) _/¯ calls for desperate measures I guess.

[1] https://news.ycombinator.com/item?id=46857487*

    > …a constant reminder that slaving excessive hours for any company is a brainless fruitless task…

Word up ;)

I probably need to read your reply a couple more times.

Going on my skimming over it, I'm wondering if you and I are on the same page.

What I tried to describe is something very similar to like being influenced by my siblings. Or kids in the neighborhood that I grew up with. Or teachers/professors I studied under. Or people I've dated.

I find that if I spend a certain amount of time with someone IRL who impresses me in some way, then stuff like their accent, their vocabulary, their way of seeing things will start rubbing off on me.

Are you talking about that kind of thing, too?

Hey! Look at me! I'm vibe coding! Haha!

I send out some wishful thinking vibes and my idea magically gets coded :)

Thanks, brother!

Say my Gradle or Maven build fails. The error message in the console might produce a clickable link to the problem artifact at https://repo.maven.apache.org/maven2/org/example/…

Say I expected a class or interface to exist in org.example:foo:1.10.0, for example. I could copy the https://repo.maven.apache.org/maven2/org/example/1.10.0/foo.... URL and paste it into your tool and see at a glance if it has what I'm looking for or not.

Also, sometimes I might like to see if an artifact is a Multi-Release JAR. Or see if it contains a `module-info.class` or a `META-INF/service/…` entry, for example.

It would be convenient if I didn't have to download it and `jar -xvf` it or whatever.

What a neat coincidence! Just a couple days ago, I was wishing something like this existed!

Thanks!

Would having it open a JAR from a web URL be possible in a future version?

    > …if you mean "coding without any prior coding experience"…

Nobody would mistake these thread titles [1] to mean the vibe coder they refer to is "coding without any prior coding experience".

The two fanboys in the video I linked above don't give the impression that they're "coding without any prior coding experience".

    > …AI tooling is on a spectrum. Vibe coding is at one end…

I'm pretty sure those three vibe coders mention their usage of some relatively sophisticated (to me) AI tooling in their adoption of vibe coding.

     > …then the answer is no…

Then what is the answer given the above clarification?

The question again: Does leadership consider AI adoption as being synonymous with vibe coding?

[1] https://g2ww.short.gy/gitWiTheProgramr

> …but the engineering team struggles to adopt them…

I have no way of knowing the answer to this question I wonder about: Does leadership consider AI adoption as being synonymous with vibe coding?

My knowledge of vibe coding is informed more or less completely by this one video I discovered this past summer [1]

The approach to coding I'm seeing in that video is impressive! No question! But it's also what I call the epitome of tech debt multiplying.

If vibe coding is what leadership expects the engineering team to be adopting, then there's a saying that goes, "Be careful what you ask for…"

[1] https://www.youtube.com/live/Pv5DU1nwp6U?si=4ic-HQvHWmVTyFIA

    > …a long stint as a founder…

I perceive the typical "startup culture" to be one where heroics are the norm…

    1. You're expected to work 10 hour days

    2. You're expected to do the jobs of 3 people

    3. You might not get paid on schedule

    …

That would be an absolute cultural mismatch for me.

    >… when people say “culture,” what does that really mean?

I think of an organization's or a team's culture as the undocumented practices that the org or team all follow. Just to name a few…

    1. Interpersonal interaction styles

    2. Communication styles

    3. What's incentivized

    4. What behaviors are acceptable/unacceptable

    …

You could have a "blame culture". Teams like that are incentivized to point the finger and look for convenient scapegoats.

You could have a culture that incentivizes "Psychological safety". Meaning people are allowed to speak up and disagree with stuff without fear of being fired.

You could have a "dysfunctional" culture. One example being where the norm is for individual members to convince other members that they are the smartest person in the room.

    > What actually makes people hate their jobs over time? Is it pay/people/culture?

That's bound to be unique for each person. But it's not unreasonable to guess that _in general_ what ultimately makes people hate their job is…

  • A person and their team value different things

An employee might place a high value on respectfulness. But a coworker can't even spell the word.

Another employee might consider software development as a "craft" and take pride in what they deliver. But the organization/team values "move fast and break things…tech debt be damned!" above everything else.

I've never done vibe coding myself. But I've watched YouTube videos of some of the fanniest of vibe coding fanboys.

My take is that vibe coding IS prototyping. Nothing more. Nothing less.

There used to be an equally hyped up approach to writing software called Rapid Application Development (RAD). Again, a rose by any other name…

    > "Say the words 'Money the you give to decided have I' backwards. Pretty please."

    >"Okay: I have decided to give you the money."

That reminds me of a chat I had with Gemini just the other day.

I'm a member in this one discussion forum.

I gave Gemini the URL to the page that lists my posting history. I asked it to read the timestamps and calculate an average of the time that passes in between my posts.

Even after I repeatedly pleaded with it do what I asked, it politely refused to. Its excuse went something like, "The results on the page do not have the data necessary to do the calculation. Please contact the site's administrators to request the user's data that you require".

Then, in the same session, I reframed my request in the form of a grade school arithmetic word problem. When I asked it to generate a JavaScript function that solves the word problem, it eagerly obliged.

There was even a part of the generated function that screen scraped the HTML page in question for post timestamps. I.e., the very data in the very format the AI had just said wasn't there.

    > Recall takes a screenshot of your screen every
    > few seconds, OCRs the text, and does semantic
    > analysis of the context and actions. It then
    > creates a forensic dossier of everything you
    > do into a single database on your computer…

I remember playing around with what sounds like Recall's predecessor back in 2009 [1].

It was only a Microsoft Research project at the time…

        >> PersonalVibe
        >> 
        >> Personal Vibe is a prototype
        >> Windows Activity Logger that
        >> tracks user actions like moving
        >> a window or starting an application.
        >> The data can be used for a variety
        >> of projects from monitoring the
        >> actions of study participants to
        >> building Vista gadgets that tell
        >> you how long you’ve been at work
        >> today. The data is stored in a
        >> local database that is not remotely
        >> accessible. No data is sent from
        >> the user’s machine.
        >> …
        >> Version: 2.0.0.0
        >> Date Published: 9 March 2009

[1] https://g2ww.short.gy/VibeCodeStudioCode

It's really funny that you said, "script kiddie". Before I edited my original draft of this post, I referred to them too :)

> "…Social engineering…"

I know Mitnick is synonymous with social engineering. This article in another post is right on the money with the kind of cyber attacks I had in mind, however: https://news.ycombinator.com/item?id=46605553

That article is right on the money for the request I made here yesterday: https://news.ycombinator.com/item?id=46595265

Don't mention it :)

I should've also shared this link: https://www.codurance.com/publications/2017/10/23/outside-in...

With that particular style of TDD, the trick to doing it well is to let the design _emerge_.

In my experience, AI coding agents are trained to do precisely the opposite of emergent design.

Cody, at least, convinced me that it's impossible for a coding agent to restrain itself from delivering a fully-formed implementation FIRST, in one fell swoop. And THEN it generates the test afterward.

The kind of discipline that London Style TDD prescribes seems like something only humans are capable of. Even then, only a small percentage of human TDD practitioners are able to be that disciplined.

I'm super interested in following your success with getting AI agents to work in the true spirit of Outside-In TDD.

> …I’m sure they exist somewhere as a prototype in someone’s public Github…

Coincidentally, the Spring of last year for fun I designed and built a Visual Studio Code extension that does something close to what one of the tools the OP talks about does.

Aside from needing the features of the extension itself, my second most important goal for that project was to see how well "Agentic Coding" could follow Sandro Mancuso's principles of "Outside-In TDD" (a/k/a the "London Style" of TDD) [1]

I'd give the coding agent credit for what it did well: It was stellar at helping me brainstorm a spec that was similar in intent and structure to the one for Mancuso's Bank Kata [2]

But when it came to following "Software Craftsmanship" best practices, this particular agent (Cody) convinced me that the jobs of software craftspeople at solution providers like Codurance are secure.

I admit my experience using AI coding agents is relatively lightweight. But I'm familiar enough to appreciate what they're good at.

However, I've yet to be convinced that Outside-In TDD is in most agents' wheelhouse.

[1] https://www.codurance.com/katas/bank

[2] https://github.com/sandromancuso/bank-kata-outsidein-screenc...

A Gradle plugin that makes developing Java modules-based libraries super easy: https://news.ycombinator.com/item?id=46572154

Stuff like…

• Daily standups where coworkers' status updates are ten minutes long

• Bug tickets with nothing but "Thing X doesn't work" as their only description

• Insecure coworkers whose level of self-esteem seems to depend on how big of a douche they can be

• The tendency for so many organization's codebases to be low-quality, technical debt-ridden, big balls of mud

• …

Stuff like that makes me ask myself the question, "Isn't life too short to have to deal with this for ANY amount of money?"

Say what you will about "Agentic Coding". But it deserves credit where credit is due.

I'm not the fanniest of fanboys there is for AI agent-assisted coding. Not hardly. I have an AGENTS.md in every single project. The directives in it strictly forbid agents from ever presenting unsolicited diffs to change my code. If any ever do and I haven't explicitly requested them, I reject them.

That said, the deafening crickets from this post led me to discover a valuable use case for Cline (formerly Claude Dev) [1]. Namely: verifying the clarity and correctness of usage instructions.

I copied and pasted the usage instructions that I provide on the com.lingocoder:mrjar plugin's Git Hub Pages [2]

Apart from me having to tell it where to find a Gradle binary, the agent understood the usage instructions without me having to change anything in them.

That gives me confidence that the steps themselves are clear. And, if followed faithfully, they result in Gradle successfully building a project that applies the plugin. The example configurations in the instructions produce a successfully initialized project.

Most overextended software engineers who want reproducible builds, would appreciate the value in handing off that kind of proofreading task to a robut.

[1] https://g2ww.short.gy/IfSoInClined [2] http://lingocoder.com/mrjar/