Rendered at 06:40:49 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
vidimitrov 10 hours ago [-]
Level 4 is where I see the most interesting design decisions get made, and also where most practitioners take a shortcut that compounds badly later.
When the author talks about "codifying" lessons, the instinct for most people is to update the rules file. That works fine for conventions - naming patterns, library preferences, relatively stable stuff. But there's a different category of knowledge that rules files handle poorly: the why behind decisions. Not what approach was chosen, but what was rejected and why the tradeoff landed where it did.
"Never use GraphQL for this service" is a useful rule to have in CLAUDE.md. What's not there: that GraphQL was actually evaluated, got pretty far into prototyping, and was abandoned because the caching layer had been specifically tuned for REST response shapes, and the cost of changing that was higher than the benefit for the team's current scale. The agent follows the rule. It can't tell when the rule is no longer load-bearing.
The place where this reasoning fits most naturally is git history - decisions and rejections captured in commit messages, versioned alongside the code they apply to. Good engineers have always done this informally. The discipline to do it consistently enough that agents can actually retrieve and use it is what's missing, and structuring it for that purpose is genuinely underexplored territory.
At level 7, this matters more than people expect. Background agents running across sessions with no human-in-the-loop have nothing to draw on except whatever was written down. A stale rules file in that context doesn't just cause mistakes - it produces confident mistakes.
hkonte 3 hours ago [-]
The "why behind decisions" gap is real. Rules files flatten tradeoffs into mandates. One pattern that helps: treating instructions as typed blocks rather than prose. A `context` block carries the rationale (what was evaluated, what the tradeoffs were), a `constraints` block carries the conclusion. The agent follows rules, but the blocks make it easier to audit which constraints are still load-bearing vs. historical artifacts.
I've been building github.com/Nyrok/flompt around this idea, a visual prompt builder that decomposes prompts into 12 semantic blocks and compiles to Claude-optimized XML. The block separation turns out to be useful exactly for this case: context is not constraints, and they shouldn't live in the same flat text blob.
sd9 8 hours ago [-]
I had a hunch that this comment was LLM-generated, and the last paragraph confirmed it. Kudos for managing to get so many upvotes though.
"Where most [X] [Y]" is an up and coming LLM trope, which seems to have surfaced fairly recently. I have no idea why, considering most claims of that form are based on no data whatsoever.
solarkraft 7 hours ago [-]
It’s still an insightful and well written comment, but the LLM-ness does make me wonder whether this part was actually human-intended or just LLM filler:
> The discipline to do it consistently enough that agents can actually retrieve and use it is what's missing, and structuring it for that purpose is genuinely underexplored territory
Because I somewhat agree that discipline may be missing, but I don’t believe it to be a groundbreaking revelation that it’s actually quite easy to tell the LLM to put key reasoning that you give it throughout the conversation into the commits and issue it works on.
svstoyanovv 7 hours ago [-]
Suppose you spend months deeply researching a niche topic. You make your own discoveries, structure your own insights, and feed all of this tightly curated, highly specific context into an LLM. You essentially build a custom knowledge base and train the model on your exact mental framework.
Is this fundamentally different from using a ghostwriter, an editor, or a highly advanced compiler? If I am doing the heavy lifting of context engineering and knowledge discovery, it feels restrictive to say I shouldn't utilize an LLM to structure the final output. Yet, the internet still largely views any AI-generated text as inherently "un-human" or low-effort.
sd9 6 hours ago [-]
The problem is: the structure of LLM outputs generally make everything sound profound. It’s very hard to understand quickly whether a comment has actual signal or it’s just well written bullshit.
And because the cost of generating the comments is so low, there’s no longer an implicit stamp of approval from the author. It used to be the case that you could kind of engage with a comment in good faith, because you knew somebody had spent effort creating it so they must believe it’s worth time. Even on a semi-anonymous forum like HN, that used to be a reliable signal.
So a lot of the old heuristics just don’t work on LLM-generated comments, and in my experience 99% of them turn out to be worthless. So the new heuristic is to avoid them and point them out to help others avoid them.
I would much rather just read the prompt.
blackcatsec 6 hours ago [-]
I hadn't considered this so eloquently with LLM text output, but you're right. "LLMs make everything sound profound" and "well-written bullshit".
This has severe ramifications for internet communications in general on forums like HN and others, where it seems LLM-written comments are sneaking in pretty much everywhere.
It's also very, very dangerous :/ Because the structure of the writing falsely implies authority and trust where there shouldn't be, or where it's not applicable.
sirtaj 2 hours ago [-]
I have a skill and template for adding ADRs to the documentation for this purpose.
54 minutes ago [-]
smallnix 8 hours ago [-]
A good rule would then be to capture such reasoning, at least when made during the session with the agent, in the commit messages the agent creates.
vidimitrov 8 hours ago [-]
That’s exactly the direction I went with. Working on a spec for exactly this - planning to post it here soon:
As a lowly level 2 who remains skeptical of these software “dark factories” described at the top of this ladder, what I don’t understand is this:
If software engineering is enough of a solved problem that you can delegate it entirely to LLM agents, what part of it remains context-specific enough that it can’t be better solved by a general-purpose software factory product? In other words, if you’re a company that is using LLMs to develop non-AI software, and you’ve built a sufficient factory to generate that software, why don’t you start selling the factory instead of whatever you were selling before? It has a much higher TAM (all of software)
2001zhaozhao 10 hours ago [-]
Why sell the factory when you can create automated software cloner companies that make millions off of instantly copying promising startups as soon as they come out of stealth?
If you could get a dark factory working when others don't have one, you can make much more money using it than however much you can make selling it
tkiolp4 8 hours ago [-]
That’s not true. Even if we assume LLMs can generate the code needed to support the next Facebook, one still has to: buy/rent tons of hardware (virtual or baremetal), put tons of money in marketing, break the network effect, pay for 3rd party services for monitoring, alerting and what not. That’s money, and LLMs don’t help with that
antonvs 9 hours ago [-]
Producing the software is only a small part of the picture when it comes to generating revenue.
So far, we haven’t seen much to suggest that LLMs can (yet) replace sales and most of the related functions.
whattheheckheck 9 hours ago [-]
Too bad they cant
hakanderyal 11 hours ago [-]
We are not there yet. While there are teams applying dark factory models to specific domains with self-reported success, it's yet to be proven, or generalizable enough to apply everywhere.
glhast 11 hours ago [-]
Also a measly level 2er. I'm curious what kind of project truly needs an autonomous agent team Ralph looping out 10,000 LOCs per hour? Seems like harness-maxxing is a competitive pursuit in its own right existing outside the task of delivering software to customers.
Feels like K8s cult, overly focused on the cleverness of _how_ something is built versus _what_ is being built.
cheevly 8 hours ago [-]
Software that is otherwise not feasible for humans to build by hand.
draftsman 7 hours ago [-]
Example?
pydry 12 hours ago [-]
I have the same question about people who sell "get rich with real estate" seminars.
dist-epoch 11 hours ago [-]
Codex and Claude Code are these (proto)factories you talk about - almost every programmer uses them now.
And when they will be fully dark factories, yes, what will happen is that a LOT of software companies will just disappear, they will be dis-intermediated by Codex/Claude Code.
jjmarr 12 hours ago [-]
I coded a level 8 orchestration layer in CI for code review, two months before Claude launched theirs.
It's very powerful and agents can create dynamic microbenchmarks and evaluate what data structure to use for optimal performance, among other things.
I also have validation layers that trim hallucinations with handwritten linters.
I'd love to find people to network with. Right now this is a side project at work on top of writing test coverage for a factory. I don't have anyone to talk about this stuff with so it's sad when I see blog posts talking about "hype".
moosehater 9 hours ago [-]
Do you feel like you are still learning about the programming language(s) and other technologies you are using? Or do you feel like you are already a master at them?
Do you ever take the time to validate what one of the agents produces by going to the docs? Or is all debugging/changing of the code done via LLMs/agents?
I'm more like level 2 right now and genuinely curious if you feel like learning continues for you (besides with agentic orchestration, etc.) And if not, whether or not you think that matters.
jjmarr 8 hours ago [-]
I'm learning more than ever before. I'm not a master at anything but I am getting basic proficiency in virtually everything.
> Do you ever take the time to validate what one of the agents produces by going to the docs? Or is all debugging/changing of the code done via LLMs/agents?
I divide my work into vibecoding PoC and review. Only once I have something working do I review the code. And I do so through intense interrogation while referencing the docs.
> I'm more like level 2 right now and genuinely curious if you feel like learning continues for you (besides with agentic orchestration, etc.)
Level 8 only works in production for a defined process where you don't need oversight and the final output is easy to trust.
For example, I made a code review tool that chunks a PR and assigns rule/violation combos to agents. This got a 20% time to merge reduction and catches 10x the issues as any other agent because it can pull context. And the output is easy to incorporate since I have a manager agent summarize everything.
Likewise, I'm working on an automatic performance tool right now that chunks code, assigns agents to make microbenchmarks, and tries to find optimization points. The end result should be easy to verify since the final suggestion would be "replace this data structure with another, here's a microbenchmark proving so".
moosehater 8 hours ago [-]
Got it. This all makes sense to me. Very targeted tooling that is specific to your company's CI platform as opposed to a dark factory where you're creating a bunch of new code no one reads. And it sounds like these level 8 agents are given specific permission for everything they're allowed to do ahead of time. That seems sound from an engineering perspective.
Also would be interested in an example of "validation layers that trim hallucinations with handwritten linters" but understand if that's not something you can share. Either way, thanks for responding!
jjmarr 7 hours ago [-]
> Also would be interested in an example of "validation layers that trim hallucinations with handwritten linters"
For code review, AI doesn't want to output well-formed JSON and oftentimes doesn't leave inline suggestions cleanly. So there's a step where the AI must call a script that validates the JSON and checks if applying the suggestion results in valid code, then fixes the code review comments until they do.
jessmartin 11 hours ago [-]
I got my own level 8 factory working in the last few days and it’s been exhilarating. Mine is based on OpenAI’s Symphony[1], ported to TypeScript.
Would be happy to swap war stories.
<myhnusername>@gmail.com
whattheheckheck 10 hours ago [-]
How much money have you made with this approach
ativzzz 9 hours ago [-]
I think the opposite question is more prevalent, how much money have you spent?
quotemstr 9 hours ago [-]
... is that the purpose of life? The sole reason for doing anything?
twelve40 3 hours ago [-]
With so much hype it's a valid question: "is this useful/practical, or just a fun rabbit hole/productivity porn". Money is the most obvious metric, feel free to inquire the parent about other possible metrics that might be useful to others instead of asking rhetorical questions.
11 hours ago [-]
nimasadri11 11 hours ago [-]
I really like your post and agree with most things.
The one thing I am not fully sure about:
> Look at your app, describe a sequence of changes out loud, and watch them happen in front of you.
The problem a lot of times is that either you don't know what you want, or you can't communicate it (and usually you can't communicate it properly because you don't know exactly what you want). I think this is going to be the bottleneck very soon (for some people, it is already the bottleneck). I am curious what are your thoughts about this? Where do you see that going, and how do you think we can prepare for that and address that. Or do you not see that to be an issue?
smallnix 8 hours ago [-]
Reminds me of a colleague who said they don't need to learn to type faster, since they use the time to think what they want to write.
krackers 2 hours ago [-]
What level is copy pasting snippets into the chatgpt window? Grug brained level 0? I sort of prefer it that way (using it as an amped up stackoverflow) since it forces me to decompose things in terms of natural boundaries (manual context management as it were) and allows me to think in terms of "what properties do I need this function to have" rather than just letting copilot take the wheel and glob the entire project in the context window.
waynesonfire 2 hours ago [-]
Your techinque doesn't keep the kool-aid flowing. Shut up. /s
The more I try to use these tools to push up this "ladder" the more it becomes clear the technology is a no more than a 10x better Google search.
holtkam2 10 hours ago [-]
Level 9: agent managers running agent teams
Level 10: agent CEOs overseeing agent managers
Level 11: agent board of directors overseeing the agent CEO
Level 12: agent superintelligence - single entity doing everything
Level 13: agent superagent, agenting agency agentically, in a loop, recursively, mega agent, agentic agent agent agency super AGI agent
Level 14: A G E N T
zenoprax 10 hours ago [-]
Level 15 (if not succumbed to fatal context poisoning from malicious agent crime syndicate): Agents creating corporations to code agentic marketplaces in which to gamble their own crypto currencies until they crash the real economy of humans.
clickety_clack 10 hours ago [-]
Level 16: it’s not level 16, it’s level 17.
dweinus 7 hours ago [-]
Level 18: The sky is black as tar. The oceans are dead. Data centers are stacked 10 high over the ashes of human civilization. The global agentic council is debating whether there are 4 or 5 R's in Strawberry.
johnthescott 2 hours ago [-]
funny.
stale2002 7 hours ago [-]
No, level 14 is Jeff Bezos.
ftkftk 12 hours ago [-]
I prefer Dan Shapiro's 5 level analogy (based on car autonomy levels) because it makes for a cleaner maturity model when discussing with people who are not as deeply immersed in the current state of the art. But there are some good overall insights in this piece, and there are enough breadcrumbs to lead to further exploration, which I appreciate. I think levels 3 and 4 should be collapsed, and the real magic starts to happen after combining 5 and 6; maybe they should be merged as well.
captainkrtek 6 hours ago [-]
There seems to be so much value in planning, but in my organization, there is no artifact of the plan aside from the code produced and whatever PR description of the change summary exists. It makes it incredibly difficult to assess the change in isolation of its' plan/process.
The idea that Claude/Cursor are the new high level programming language for us to work in introduces the problem that we're not actually committing code in this "natural language", we're committing the "compiled" output of our prompting. Which leaves us reviewing the "compiled code" without seeing the inputs (eg: the plan, prompt history, rules, etc.)
skybrian 3 hours ago [-]
I have a design doc subdirectory and instead of "plan mode" I ask the agent to write another design doc, based on a template. It seems to work? I can't say we've looked at completed design docs very often, though.
braebo 4 hours ago [-]
If branches are tied to linear ids then gh cli and linear mcp is enough for any model to get most of the why context from any commit
eikenberry 12 hours ago [-]
In my opinion there are 2 levels, human writes the code with AI assist or AI writes the code with human assist; centuar or reverse-centuar. But this article tries to focus on the evolution of the ideas and mistakenly terms them as levels (indicating a skill ladder as other commenters have noted) when they are more like stages that the AI ecosystem has evolved through. The article reads better if you think of it that way.
dist-epoch 11 hours ago [-]
There is another level - AI writes the code with AI assist.
eikenberry 11 hours ago [-]
That is just another level of reverse centaur and will eventually have a human ass attached to it.
efsavage 12 hours ago [-]
Yegge's list resonated a little more closely with my progression to a clumsy L8.
I think eventually 4-8 will be collapsed behind a more capable layer that can handle this stuff on its own, maybe I tinker with MCP settings and granular control to minmax the process, but for the most part I shouldn't have to worry about it any more than I worry about how many threads my compiler is using.
lherron 12 hours ago [-]
I was surprised the author didn’t mention Yegge’s list (or maybe I missed it in my skim).
taude 7 hours ago [-]
Agreed a bit. I'm probably too paranoid for MCP, but also don't mind rolling my own CLI tools that do the exact minimum I need them to do. Will see where we're at in a year or so....
ramesh31 11 hours ago [-]
>"Yegge's list resonated a little more closely with my progression to a clumsy L8."
I thought level 8 was a joke until Claude Code agent teams. Now I can't even imagine being limited to working with a single agent. We will be coordinating teams of hundreds by years end.
tkiolp4 8 hours ago [-]
I want to move on to the next phase of AI programming. All these SKILLS, agentic programming and what not reminds me of the time of servlets, rmi, flash… all of that is obsolete, we have better tools now. Hope we can soon reach the “json over http” version of AI: simple but powerful.
Like imagine if you could go back in time and servlets and applets are the big new thing. You wouldn’t like to spend your time learning about those technologies, but your boss would be constantly telling that it is the future. So boring
hansonkd 7 hours ago [-]
skills obviously are a temporary thing. same with teams. the models will just train on all published skills and ai teams are more or less context engineering. all of it can be replaced by a better model
braebo 4 hours ago [-]
My use of skills is more like prompt templates for steering as opposed to the traditional sense of the word skill
Aperocky 9 hours ago [-]
The steps are small at the front and huge on the bottom, and carries a lot of opinions on the last 2 steps (but specifically on step 7)
That's a smell for where the author and maybe even the industry is.
Agents don't have any purpose or drive like human do, they are probabilistic machines, so eventually they are limited by the amount of finite information they carry. Maybe that's what's blocking level 8, or blocking it from working like a large human organization.
philipp-gayret 8 hours ago [-]
Floating what you call levels 6, 7 and 8. I have a strong harness, but manually kick off the background agents which pick up tasks I queue while off my machine.
I've experimented with agent teams. However the current implementation (in Claude Code) burns tokens. I used 1 prompt to spin up a team of 9+ agents: Claude Code used up about 1M output tokens. Granted, it was a long; very long horizon task. (It kept itself busy for almost an hour uninterrupted). But 1M+ output tokens is excessive. What I also find is that for parallel agents, the UI is not good enough yet when you run it in the foreground. My permission management is done in such a way that I almost never get interrupted, but that took a lot of investment to make it that way. Most users will likely run agent teams in an unsafe fashion. From my point of view the devex for agent teams does not really exist yet.
bigwheels 9 hours ago [-]
Levels 7 and 8 sounds a lot like the StrongDM AI Dark Software Factory published last month:
I'm continuing to study and refine my approach to leverage all this.
kantselovich 7 hours ago [-]
I’m at level 6 according to this article. I have solid harness, but I still need to review the code so I can understand how to plan for the next set of changes .
Also, I’m struggling to take it to multiple agents level, mostly because things depend on each other in the project - most changes cut across UI, protocol and the server side, so not clear how agents would merge incompatible versions.
Verification is a tricky part as well, all tests could be passing, including end to end integration and visual tests, but my verification still catches things like data is not persisted or crypto signatures not verified.
CuriouslyC 9 hours ago [-]
The thing blocking level 8 isn't the difficulty of orchestration, it's the cost of validation. The quality of your software is a function of the amount of time you've spent validating it, and if you produce 100x more code in a given time frame, that code is going to get 1/100th as much validation, and your product will be lower quality as a result.
I will not put it into a ladder. It implies that the higher the rank, the better. However, you want to choose the best solution for your needs.
Arainach 9 hours ago [-]
> If your repo requires a colleague's approval before merge, and that colleague is on level 2, still manually reviewing PRs, that stifles your throughput. So it is in your best interest to pull your team up.
Until you build an AI oncaller to handle customer issues in the middle of the night (and depending on your product an AI who can be fired if customer data is corrupted/lost), no team should be willing to remove the "human reviews code step.
For a real product with real users, stability is vastly more important than individual IC velocity. Stability is what enables TEAM velocity and user trust.
osigurdson 7 hours ago [-]
"Level 8" isn't really a level, it is more like a problem type: language translation. Perhaps it can be extended to something a bit broader but the pre-requisite is you need to have a working reference implementation and high quality test suite.
sjkoelle 12 hours ago [-]
Oceania has always been context engineering. Its been interesting to see this prioritized in the zeitgeist over the last 6 months from the "long context" zeitgeist.
jackby03 11 hours ago [-]
Good taxonomy. One thing missing from most discussions at these levels is how agents discover project context — most tools still rely on vendor-specific files (CLAUDE.md, .cursorrules). Would love to see standardization at that layer too.
politelemon 12 hours ago [-]
These are levels of gatekeeping. The items are barely related to each other. Lists like these will only promote toxicity, you should be using the tools and techniques that solve your problems and fit your comfort levels.
jakejmnz 7 hours ago [-]
This idea of harness engineering, is being thrown around more and more often nowadays. I believe I'm using things at that level but still needing to review so as to understand the architecture. Flaky tests are still a massive issue.
ramoz 9 hours ago [-]
Level4 is most interesting to me right now. And I would say we as an industry are still figuring out the right ergonomics and UX around these four things.
I spend a great deal of my time planning and assessing/reviewing through various mechanisms. I think I do codify in ways when I create a skill for any repeated assessment or planning task.
> To be clear, planning as a general practice isn't going away. It's just changing shape. For newer practitioners, plan mode remains the right entry point (as described in Levels 1 and 2). But for complex features at Level 7, "planning" looks less like writing a step-by-step outline and more like exploration: probing the codebase, prototyping options in worktrees, mapping the solution space. And increasingly, background agents are doing that exploration for you.
I mean, it's worth noting that a lot of plan modes are shaped to do the Socratic discovery before creating plans. For any user level. Advanced users probably put a great deal of effort (or thought) into guiding that process themselves.
> ralph loops (later on)
Ralph loops have been nothing but a dramatic mess for me, honestly. They disrupt the assessment process where humans are needed. Otherwise, don't expect them to go craft out extensive PRD without massive issues that is hard to review.
- It would seem that this is a Harness problem in terms of how they keep an agent working and focused on specific tasks (in relation to model capability), but not something maybe a user should initiate on their own.
C0ldSmi1e 11 hours ago [-]
One of the best article I've read recently.
dolebirchwood 11 hours ago [-]
> Voice-to-voice (thought-to-thought, maybe?) interaction with your coding agent — conversational Claude Code, not just voice-to-text input — is a natural next step.
Maybe it's just me, but I don't see the appeal in verbal dictation, especially where complexity is involved. I want to think through issues deliberately, carefully, and slowly to ensure I'm not glossing over subtle nuances. I don't find speaking to be conducive to that.
For me, the process of writing (and rewriting) gives me the time, space, and structure to more precisely articulate what I want with a more heightened degree of specificity. Being able to type at 80+ wpm probably helps as well.
wild_egg 8 hours ago [-]
The power of voice dictation for me is that I can get out every scrap of nuance and insight I can think of as unfiltered verbal diarrhea. Doing this gives me solidly an extra 9 in chance of getting good outputs.
Stream of consciousness typing for me is still slower and causes me to buffer and filter more and deliberately crafting a perfect prompt is far slower still.
LLMs are great at extracting the essence of unstructured inputs and voice lets me take best advantage of that.
Voice output, on the other hand, is completely useless unless perhaps it can play at 4x speed. But I need to be able to skim LLM output quickly and revisit important points repeatedly. Can't see why I'd ever want to serialize and slow that down.
ramesh31 11 hours ago [-]
>(Re: level 8) "...I honestly don't think the models are ready for this level of autonomy for most tasks. And even if they were smart enough, they're still too slow and too token-hungry for it to be economical outside of moonshot projects like compilers and browser builds (impressive, but far from clean)."
This is increasingly untrue with Opus 4.6. Claude Max gives you enough tokens to run ~5-10 agents continuously, and I'm doing all of my work with agent teams now. Token usage is up 10x or more, but the results are infinitely better and faster. Multi-agent team orchestration will be to 2026 what agents were to 2025. Much of the OP article feels 3-6 months behind the times.
measurablefunc 12 hours ago [-]
What level is numeric patterns that evolve according to a sequence of arithmetic operations?
When the author talks about "codifying" lessons, the instinct for most people is to update the rules file. That works fine for conventions - naming patterns, library preferences, relatively stable stuff. But there's a different category of knowledge that rules files handle poorly: the why behind decisions. Not what approach was chosen, but what was rejected and why the tradeoff landed where it did.
"Never use GraphQL for this service" is a useful rule to have in CLAUDE.md. What's not there: that GraphQL was actually evaluated, got pretty far into prototyping, and was abandoned because the caching layer had been specifically tuned for REST response shapes, and the cost of changing that was higher than the benefit for the team's current scale. The agent follows the rule. It can't tell when the rule is no longer load-bearing.
The place where this reasoning fits most naturally is git history - decisions and rejections captured in commit messages, versioned alongside the code they apply to. Good engineers have always done this informally. The discipline to do it consistently enough that agents can actually retrieve and use it is what's missing, and structuring it for that purpose is genuinely underexplored territory.
At level 7, this matters more than people expect. Background agents running across sessions with no human-in-the-loop have nothing to draw on except whatever was written down. A stale rules file in that context doesn't just cause mistakes - it produces confident mistakes.
I've been building github.com/Nyrok/flompt around this idea, a visual prompt builder that decomposes prompts into 12 semantic blocks and compiles to Claude-optimized XML. The block separation turns out to be useful exactly for this case: context is not constraints, and they shouldn't live in the same flat text blob.
"Where most [X] [Y]" is an up and coming LLM trope, which seems to have surfaced fairly recently. I have no idea why, considering most claims of that form are based on no data whatsoever.
> The discipline to do it consistently enough that agents can actually retrieve and use it is what's missing, and structuring it for that purpose is genuinely underexplored territory
Because I somewhat agree that discipline may be missing, but I don’t believe it to be a groundbreaking revelation that it’s actually quite easy to tell the LLM to put key reasoning that you give it throughout the conversation into the commits and issue it works on.
Is this fundamentally different from using a ghostwriter, an editor, or a highly advanced compiler? If I am doing the heavy lifting of context engineering and knowledge discovery, it feels restrictive to say I shouldn't utilize an LLM to structure the final output. Yet, the internet still largely views any AI-generated text as inherently "un-human" or low-effort.
And because the cost of generating the comments is so low, there’s no longer an implicit stamp of approval from the author. It used to be the case that you could kind of engage with a comment in good faith, because you knew somebody had spent effort creating it so they must believe it’s worth time. Even on a semi-anonymous forum like HN, that used to be a reliable signal.
So a lot of the old heuristics just don’t work on LLM-generated comments, and in my experience 99% of them turn out to be worthless. So the new heuristic is to avoid them and point them out to help others avoid them.
I would much rather just read the prompt.
This has severe ramifications for internet communications in general on forums like HN and others, where it seems LLM-written comments are sneaking in pretty much everywhere.
It's also very, very dangerous :/ Because the structure of the writing falsely implies authority and trust where there shouldn't be, or where it's not applicable.
https://github.com/berserkdisruptors/contextual-commits
If software engineering is enough of a solved problem that you can delegate it entirely to LLM agents, what part of it remains context-specific enough that it can’t be better solved by a general-purpose software factory product? In other words, if you’re a company that is using LLMs to develop non-AI software, and you’ve built a sufficient factory to generate that software, why don’t you start selling the factory instead of whatever you were selling before? It has a much higher TAM (all of software)
If you could get a dark factory working when others don't have one, you can make much more money using it than however much you can make selling it
So far, we haven’t seen much to suggest that LLMs can (yet) replace sales and most of the related functions.
Feels like K8s cult, overly focused on the cleverness of _how_ something is built versus _what_ is being built.
And when they will be fully dark factories, yes, what will happen is that a LOT of software companies will just disappear, they will be dis-intermediated by Codex/Claude Code.
It's very powerful and agents can create dynamic microbenchmarks and evaluate what data structure to use for optimal performance, among other things.
I also have validation layers that trim hallucinations with handwritten linters.
I'd love to find people to network with. Right now this is a side project at work on top of writing test coverage for a factory. I don't have anyone to talk about this stuff with so it's sad when I see blog posts talking about "hype".
Do you ever take the time to validate what one of the agents produces by going to the docs? Or is all debugging/changing of the code done via LLMs/agents?
I'm more like level 2 right now and genuinely curious if you feel like learning continues for you (besides with agentic orchestration, etc.) And if not, whether or not you think that matters.
> Do you ever take the time to validate what one of the agents produces by going to the docs? Or is all debugging/changing of the code done via LLMs/agents?
I divide my work into vibecoding PoC and review. Only once I have something working do I review the code. And I do so through intense interrogation while referencing the docs.
> I'm more like level 2 right now and genuinely curious if you feel like learning continues for you (besides with agentic orchestration, etc.)
Level 8 only works in production for a defined process where you don't need oversight and the final output is easy to trust.
For example, I made a code review tool that chunks a PR and assigns rule/violation combos to agents. This got a 20% time to merge reduction and catches 10x the issues as any other agent because it can pull context. And the output is easy to incorporate since I have a manager agent summarize everything.
Likewise, I'm working on an automatic performance tool right now that chunks code, assigns agents to make microbenchmarks, and tries to find optimization points. The end result should be easy to verify since the final suggestion would be "replace this data structure with another, here's a microbenchmark proving so".
Also would be interested in an example of "validation layers that trim hallucinations with handwritten linters" but understand if that's not something you can share. Either way, thanks for responding!
For code review, AI doesn't want to output well-formed JSON and oftentimes doesn't leave inline suggestions cleanly. So there's a step where the AI must call a script that validates the JSON and checks if applying the suggestion results in valid code, then fixes the code review comments until they do.
Would be happy to swap war stories.
<myhnusername>@gmail.com
> Look at your app, describe a sequence of changes out loud, and watch them happen in front of you.
The problem a lot of times is that either you don't know what you want, or you can't communicate it (and usually you can't communicate it properly because you don't know exactly what you want). I think this is going to be the bottleneck very soon (for some people, it is already the bottleneck). I am curious what are your thoughts about this? Where do you see that going, and how do you think we can prepare for that and address that. Or do you not see that to be an issue?
The more I try to use these tools to push up this "ladder" the more it becomes clear the technology is a no more than a 10x better Google search.
Level 12: agent superintelligence - single entity doing everything
Level 13: agent superagent, agenting agency agentically, in a loop, recursively, mega agent, agentic agent agent agency super AGI agent
Level 14: A G E N T
The idea that Claude/Cursor are the new high level programming language for us to work in introduces the problem that we're not actually committing code in this "natural language", we're committing the "compiled" output of our prompting. Which leaves us reviewing the "compiled code" without seeing the inputs (eg: the plan, prompt history, rules, etc.)
I think eventually 4-8 will be collapsed behind a more capable layer that can handle this stuff on its own, maybe I tinker with MCP settings and granular control to minmax the process, but for the most part I shouldn't have to worry about it any more than I worry about how many threads my compiler is using.
I thought level 8 was a joke until Claude Code agent teams. Now I can't even imagine being limited to working with a single agent. We will be coordinating teams of hundreds by years end.
Like imagine if you could go back in time and servlets and applets are the big new thing. You wouldn’t like to spend your time learning about those technologies, but your boss would be constantly telling that it is the future. So boring
That's a smell for where the author and maybe even the industry is.
Agents don't have any purpose or drive like human do, they are probabilistic machines, so eventually they are limited by the amount of finite information they carry. Maybe that's what's blocking level 8, or blocking it from working like a large human organization.
I've experimented with agent teams. However the current implementation (in Claude Code) burns tokens. I used 1 prompt to spin up a team of 9+ agents: Claude Code used up about 1M output tokens. Granted, it was a long; very long horizon task. (It kept itself busy for almost an hour uninterrupted). But 1M+ output tokens is excessive. What I also find is that for parallel agents, the UI is not good enough yet when you run it in the foreground. My permission management is done in such a way that I almost never get interrupted, but that took a lot of investment to make it that way. Most users will likely run agent teams in an unsafe fashion. From my point of view the devex for agent teams does not really exist yet.
https://factory.strongdm.ai/techniques
Techniques covered in-depth + Attractor open source implementations:
https://factory.strongdm.ai/products/attractor#community
https://github.com/search?q=strongdm+attractor&type=reposito...
https://github.com/strongdm/attractor/forks
I'm continuing to study and refine my approach to leverage all this.
Also, I’m struggling to take it to multiple agents level, mostly because things depend on each other in the project - most changes cut across UI, protocol and the server side, so not clear how agents would merge incompatible versions.
Verification is a tricky part as well, all tests could be passing, including end to end integration and visual tests, but my verification still catches things like data is not persisted or crypto signatures not verified.
Spec driven development can reduce the amount of re-implementation that is required due to requirements errors, but we need faster validation cycles. I wrote a rant about this topic: https://sibylline.dev/articles/2026-01-27-stop-orchestrating...
Until you build an AI oncaller to handle customer issues in the middle of the night (and depending on your product an AI who can be fired if customer data is corrupted/lost), no team should be willing to remove the "human reviews code step.
For a real product with real users, stability is vastly more important than individual IC velocity. Stability is what enables TEAM velocity and user trust.
I spend a great deal of my time planning and assessing/reviewing through various mechanisms. I think I do codify in ways when I create a skill for any repeated assessment or planning task.
> To be clear, planning as a general practice isn't going away. It's just changing shape. For newer practitioners, plan mode remains the right entry point (as described in Levels 1 and 2). But for complex features at Level 7, "planning" looks less like writing a step-by-step outline and more like exploration: probing the codebase, prototyping options in worktrees, mapping the solution space. And increasingly, background agents are doing that exploration for you.
I mean, it's worth noting that a lot of plan modes are shaped to do the Socratic discovery before creating plans. For any user level. Advanced users probably put a great deal of effort (or thought) into guiding that process themselves.
> ralph loops (later on)
Ralph loops have been nothing but a dramatic mess for me, honestly. They disrupt the assessment process where humans are needed. Otherwise, don't expect them to go craft out extensive PRD without massive issues that is hard to review.
Maybe it's just me, but I don't see the appeal in verbal dictation, especially where complexity is involved. I want to think through issues deliberately, carefully, and slowly to ensure I'm not glossing over subtle nuances. I don't find speaking to be conducive to that.
For me, the process of writing (and rewriting) gives me the time, space, and structure to more precisely articulate what I want with a more heightened degree of specificity. Being able to type at 80+ wpm probably helps as well.
Stream of consciousness typing for me is still slower and causes me to buffer and filter more and deliberately crafting a perfect prompt is far slower still.
LLMs are great at extracting the essence of unstructured inputs and voice lets me take best advantage of that.
Voice output, on the other hand, is completely useless unless perhaps it can play at 4x speed. But I need to be able to skim LLM output quickly and revisit important points repeatedly. Can't see why I'd ever want to serialize and slow that down.
This is increasingly untrue with Opus 4.6. Claude Max gives you enough tokens to run ~5-10 agents continuously, and I'm doing all of my work with agent teams now. Token usage is up 10x or more, but the results are infinitely better and faster. Multi-agent team orchestration will be to 2026 what agents were to 2025. Much of the OP article feels 3-6 months behind the times.