I keep wondering how people accept a nights worth of agent activity.
I feel 30 minutes of planning and 30 minutes of implementation in my solo side project's repo is too big to review. At minute 5, I may ask the AI to redo stuff even as its spitting out code.
Most of the narrative is about how AI is writing all/most code, but I’d wager that the fraction of human reviewed code is approaching zero far faster than anyone is realizing or willing to admit.
Very true. Last year I at least glanced at every line of AI generated code. Now if some AI makes a 10k line program for some one-off tasks, I run the program, glance only over the output, and move on.
Especially if you're having an LLM write non-interactive scripts to calculate complex things from large datasets, glancing at the output is not enough to know if the output is remotely accurate (unless the output is so trivial you could literally do it in your head).
Case in point: I recently asked an LLM to write a pile of code to compile historical baseball stats to test betting success against the results of my hand-written code that evolves genetic algorithms. I marveled for a little while at the unbelievable improvement in EV/ROI that this script was showing could have been achieved from certain small tweaks. I only noticed after pushing a total bet that the push registered on the output as a win - and only because I was carefully staying on top of it. A single stupid recursively operating >= instead of > had caused completely nonsensical results that looked plausible.
Imagine, like, trusting a 10k loc script to give you data for something you were going to build in the physical world, and hoping an LLM hadn't made a mistake like that.
Would depend on what AI and prompt you use ultimately. Ask it to add tests (functional, E2E and unit, maybe invent a new type too), packaging, modular code and/or whatever, and you get to 10K relatively quickly with some of the more verbose LLMs out there.
Personally it's probably the biggest struggle, trying to rein in the "spray and pray" approach LLMs typically like to take, and reducing the "patch on top of patch" syndrome too.
Calculate the engine power of a 2015 VW polo when travelling 70 mph on a flat road behind a box truck. Draw a chart of drag Vs follow distance. How significant is humidity on the result?
The assumption used to be that you respected the library enough and believed it was well reviewed and architected by the maintainer(s). But now even that's unreliable because libraries are being slopified at an unreviewable pace too.
It's weird that you think humans weren't slopifying code until LLM's came along. At least now they are implementing tests and CI and far more documentation, updating API versions, etc. OOMs above the amount they did before.
I'd also wager that far more % of code gets more coverage of review, via prompting AI to do it, than it did before.
Most PR's pass as long as they A. pass checks, B. dont introduce regressions, C. fix a bug or implement a feature. People talk about this era of humans reviewing code with nostalgia... but that never existed at scale.
I’d say the increased scrutiny has merely exposed the difference in care between the different groups in the industry. Seems to explain pretty well why both sides are equally confounded by the other’s expectations.
People keep saying this like it’s some meaningful point, but the reality is many people in different projects have a shared need for that code to work correctly, and there is a social proof involved in used open source libraries. That is why people look at downloads and dependent projects as heuristics of stability and correctness. That is not the case with (and cannot be obtained with) code authored by generative AI.
Yes it can, the code will be ran and you will have the proof that it ran well. Or it won't run well and you'll re-do it. Same as with some imported library.
A lot of that agent activity is combing over what was previously made, forcing constraints upon it so you have a reasonable expectation of what ends up on your desk for review.
For me, strong file structure helps as well. Reviewing a 3,000 line file it just created is abysmal. I wouldn't accept that from human nor machine :) Multiple files in the right places helps reduce cognitive load.
Sometimes I'll also review with the agent interactively. What is the most important file to review first, etc?
I like to stage changes into a "LGTM" pile. Then if I want changes, I'll have the agent "review unstaged changes - I want something different done here."
No one is reviewing the code. Managers don't want us to review code either. It's a bottleneck. If something goes wrong (bugs) they are fixed as they come. It's a very sad era of software engineering. If there ever was some engineering in our trade, now it's mostly gone. We are guessing around, writing "skills" files with "please, do not introduce bugs" or "you are an owner, not a renter" or similar stuff. It's just very low effort, very undeterministic. Big apps out there are going down constantly because of AI slop (e.g., Github), and we are seeing it more often as well in non-so popular systems (e.g., in my company and other saas that we use).
Product managers never cared about the code. Engineering managers don't care about code as much as they did when they were engineers. Directors couldn't care less about code. CTOs don't know what code looks like anymore. We are at the end of the chain, and somehow we always took pride of well written and maintainble code because we knew deep inside that good systems are built based on good code. But now we are jeopardizing ourselves, it's us the engineers who don't care anymore about code and with AI that problems is amplified.
I usually aim to have Claude end up with about 500 lines of code after a night of work. Most of what it's doing is experimenting with many different approaches, summarizing them, and then giving me a relatively small diff to review and modify.
This is the way to go. I usually play with relatively stable software where the improvements are either performance or very small niche features that are built on top of already existing ones. Big changes are undesirable by both the others working on it and its users.
You care about code quality. Many don’t. I had someone tell me this week that a 6000 line class was ok because it was easier for the model to understand and that’s more important than human comprehension. And I get his point but that seems like a big risk to take.
and it's wrong. a 6000 line class is not easier for a model to understand. the same things that help humans also help agents. I find myself adding linters that must pass and the agent muss fix that limit file size, function length, function complexity, how many files in a directory. a little more work for the agent, but the codebase is healthier and the agents write fewer bugs.
Parsing single file is easier than navigating a file system for an LLM. Until the models have context windows large enough to hold the entire codebase in one shot, single files will beat multiple files every time.
Yeah the multi-agent workflow just hasn't been satisfying to me. The more chats I try to run at once, the more I got lost and overwhelmed. I trust Claude to implement a plan correctly after I've reviewed it, but if I don't review all of the plans, I will miss some small detail that it misunderstood and it'll be a pain to fix later.
I'm like a 1-2 chats at a time kind of guy. I just don't see how I could keep my exact vision for the project otherwise.
Same, on top of that multi-agent workflows just cost too much to make stopping and correcting them to feel worthwhile, compared to one or two manually managed chats
Even the newest models, like GPT 5.5, only deliver what I want nine out of ten times. If I didn't catch the remaining 10% of misguided garbage by manually reviewing every change, it would add up really quickly.
I never look at code. It used to be that it quickly became unmaintainable spaghetti where the agent struggled to make any change at all, but in the past year (and with a three step plan/develop/review workflow), the quality is so good that I basically just don't look at the code any more.
It definitely has fewer bugs than a senior developer, but it really hinges on getting the plan right. 20 minutes of planning and 20 of implementation sounds about right for my workflow as well, just make sure you have GPT as a reviewer. It's very nitpicky and finds lots of bugs.
First, that this is challenging to scale across large orgs. Even if your plans produce high quality code, that isn’t true for everyone. I’m definitely struggling with slop code being collectively mailed to me for review my our 1,000 engineers that were told to use their AI subscription all at once.
I feel like we should be taking “prompt engineering” more seriously. And when people mail me code to review, it should also include the agentic workflow and plan. So that when code isn’t up to quality, and can have a discussion about the prompts used to generate it.
My second thought is related to your senior engineer comment. This isn’t surprising, because in most engineering orgs, seniority is completely unrelated to code quality. In fact, many orgs incentive the opposite: “senior” devs that push out buggy code quickly and push accountability downhill to the junior devs.
Eh, everything is challenging to scale across large orgs. Even before LLMs, the code was a huge ball of spaghetti that barely held together. Now we just get there faster.
About senior engineers, I guess that depends on the org you have experience with. My experience doesn't match yours.
So I've been in a hobby project for a few weeks -- transforming an old software modem binary to c code.
I gave it the existing modem, and had it build rigging to build test vectors. I had it specify the work in the modem. And to confirm that legacy<>legacy produced the same streams as the new code. I've also recorded test vectors vs. other modems.
I've since launched it on targeted refactoring and code reduction projects.
I am mostly not looking at the code. There's a 100KSLOC lump of code that is much cleaner than a decompilation but a fair bit dirtier than what I would write myself. It is not factored terribly. I have some hope of getting it to trim this down to 70KSLOC that then I can accept in small blocks.
It outperforms the original softmodem, hitting higher RX rates for the same line quality and using less CPU. It also has additional functionality.
So, you know, I would never have written something this large for a hobby myself. And it's cost me $200 and 20-30 minutes per day for a few weeks to get a huge functional surface that I do believe I will be able to trust at the end of the process.
That depends. When I'm working on a 1 in a million race condition in some multi-threaded code, the agent needs hours to figure out what is going on. (I would probably need weeks - I don't know as I've given up on some of these before I could point an agent at it)
I agree, but for small tasks - <20 lines that I can understand in a minute or two - perfect. Thinking about it - I have hundreds, if not thousands of tasks that I would like to do, improving pipelines, migrating from one tool to another, but never have time. The only question is - if I don't have time to do it, do I have time to prompt it?
whenever i found a guy who uses parallel overnight agents, i asked them how many users they have. Crickets.
They do not have any users. Meanwhile, i've to do code reviews and all otherwise my 12,000+ users will be pissed off if anything in their workflow breaks.
This means i really cannot release more than 1 tiny feature a day. And using parallel agents, well that's good for testing but i don't think i need to add that many features to add anything.
Lots of people are working on repetitive simple projects like the Nth website whatever or things like that, boring stuff. This LLM era is already a very big deal for these people.
Personally somehow I am working on stuff that has like 25% not trivial stuff and that is enough to have the same experience as you have.
But also lots of people just don't care about quality and they might be right with their customers/audience. In these cases when someone catches one, an agent is going to iterate on it and make it (seemingly) go away, bandage applied, who cares again. This has a market, I am sure. Lots of programmer folks are just as bad.
One thing I don't understand with all of these, is how they're handling different worktree infrastructure spin-up?
For example, if I have a webapp, I want each of the worktrees to spin up its own infrastructure, and be accessible on its own unique local url, so that I can see the changes locally for each worktree, or I can have agents automate visual checks using something like agent-browser.
Currently I use docker for my infrastructure, each service running in its own container. I have a script that has a ./app worktree create worktreename. That creates a worktree as "worktreename" and spins up all of my docker infrastructure with prefixes for things like "WORKTREENAME", and I can access all my urls at worktreename.myapp.test (or just myapp.test for the main worktree).
This is working fine for now, but it'd be cool if one of these apps was compatible with this concept so I could move over to that.
I have this issue at work, I asked CC to create a very simple bun CLI tool that can be hard to create, destroy and list worktrees.
The CLI seeds the .env file with a url, db for that worktree and I use Vercels open source package portless to spin up a dev server with unique ports so I get a url per worktree
Just use direnv? You’ll probably need to adjust the port you are hosting the local page on, but that’s just N=mod(hash based on the worktree name) and then port=default_port+N.
Tell your claude to set this up. Should do it in a single prompt
This reminds me of Vibe Kanban (https://vibekanban.com/) which I use to manage coding agents on most of my projects.
The Vibe Kanban developers unfortunately decided that they didn't see a path to profitability and have stopped investing in the project. It's open source and so you can run it locally / fork it, but it has stopped improving and there are still annoying bugs that need to be fixed (and I don't have time to maintain it personally). This makes me sad because I would be willing to pay for Vibe Kanban, but I didn't need the features their paid plan offered (in retrospect maybe I should have paid anyway).
I'll give Kanbots a go :) I'd recommend liberally copying features from Vibe Kanban. In particular the remote support and "Open in VS Code" button (which in my case opens a local VSCode client pointing to a remote VSCode server) are critical for me.
Vibe Kanban is indeed a treasure trove in terms of useful features.
I've been working for the last week or two on getting my new tool up to parity with VK with additional improvements. I've been posting some screenshots into the Vibe Kanban discord as well. Hopefully it'll be a great fit for your use case when I finally am ready to launch it.
(My tool aims for better features than VK in both the Kanban board and agent workspaces, while adding extra systems like desktop windowing, plugins, in-browser VSCode integration, and htmx-like server-rendered UI. The remote access also works differently - you host the whole thing like OpenClaw and access the remote desktop UI from the browser, rather than run a webserver on your laptop to access remote coding agents.)
> "Local-first, zero servers. Everything lives in .kanbots/ next to your repo: SQLite database, configs, worktrees. No cloud account, no telemetry, no HTTP server. This is the open-source desktop edition."
This is table-stakes for me to consider adoption of a tool like this.
If AI is agentic I would expect it takes an hour of chatting for any PM to integrate some agent Ralph loop with Jira. Jira or Trello or Linear or Basecamp all have APIs and I guess CLIs any agent can use to talk to them. No developer or SaaS should be needed to make them understand tasks are checked out when you start work and contain instructions and when you are done you move the ticket to DONE.
From their page, they say they require cloud account login for this to work, even locally which is why I decided not to try it out. Looks cool tbh. But I have quite a few tools that look cool.
The minimum required payment to play a gambling game, where the money up for grabs is called "stake". See also "raising the stakes". In context it means the minimum feature set to be considered for adoption.
I have got more frustrations than successes when I tried to run agents without supervising them. I believe the technology will get there eventually, but right now I need one IDE per agent and its cumbersome to merge the work.
There's a few apps out there that facilitate handing off to agents from kanban boards. I needed something more 'human in the loop', handing off to an agent without good visibility of the change set and opportunity to steer doesn't work for me. https://www.agentkanban.io links a taskboard with github copilot chat in vs code via our extension so we have the benefit of task management and context capture from the chat to the tasks. This gives us all the features of a top harness (vs code) and the task / project management features at the same time.
Hello folks, sharing my latest open source project, a kanban board with parallel agents. Trying to improve this with more features, I would love your contributions on this repo, with either code contributions or ideas
Nice work .. I have had my own agents running kanban on existing Jira projects, categorized by workflow, and it is a pleasure to see your project on HN today. I will for sure enjoy catching up with your work, thanks for sharing it.
Gave it a brief shot, felt a bit early on, went back to Claude. I feel like the Kanban board that would do it best would just allow easily bringing up Claude Code sessions with all user input etc.
If Jujutsu had become popular a couple of years earlier, it might have had a chance to catch on. I worry that it missed the most critical training window for AI and we may be locked in on Git forever.
This is dope! We basically built something very similar internally for our team and it's been a very natural and intuitive way to manage agents (as opposed to having a bunch of terminals to track). Not every task/conversation can be done in the background, so it's been helpful for us internally to be able to seamlessly transition between "interactive conversation" and "background job done by agents" even within a single card.
i was also building something similar. https://github.com/jaequery/planbooq. i think building your own kanban is the new age vim/emacs, so you can streamline your workflows the way you want them to be.
Personally, this is somewhat close to what I want.
I want to have a fullblown cursor instance/window for each task I have, and a central Hub that manages spawning those instances, setting up the worktrees, etc.
Cursor seems to pretty much have all the available tools there already (it can already spawn agents to their own worktrees with proper setup scripts, for example). I don't get why they don't do it and instead insist on a buggy and confusing agents experience.
Unfortunately, most attempts at this seem to assume I want a model where "1 task = 1 agent = 1 chat", whereas what I really want is "1 task = 1 worktree = 1 full IDE around it".
With the full IDE I can have multiple agents/conversations, review code thoroughly and also chip in once in a while. I can have multiple models (that I pick) in multiple chats, iterate forwards, backwards, you name it.
I really don't understand why there seems to be this idea that "parallel agents" should live in their own little restricted flow that's limited to a tiny chat interface. I want the full flow for every agent!
I was hoping cursor would do this, but they really seem to be going the direction of turning their absolutely terrible web agents UI (where you can't even CHANGE THE MODEL!!!!) into a desktop thing. Sad, as I've been an Ultra paying customer and might have to leave soon with the direction they're heading.
> I want to have a fullblown cursor instance/window for each task I have, and a central Hub that manages spawning those instances, setting up the worktrees, etc.
I am working on exactly this interface for my new tool called Kotkit. You start with kanban board management of workspaces. Each workspace (worktree on one/multiple repos) is a feature-rich IDE interface in a remote-capable in-browser desktop. You can spawn multiple agents with a good UI wrapper and full auditable logs, solve worktree rebase/merge with 1-click AI features, and there is also an embedded VSCode to solve edge cases. It also supports very deep plugin integration like IntelliJ.
Currently dogfooding it on my own projects and will be released sometime soon.
The parallel agents concept is interesting. How does it handle state sync between agents when they're modifying the same board? Or is there built-in conflict resolution?
one of your pages return 404 /comparison... but cool project! I guess we're just still not there to let agents run without supervision. At least for me.
Just post the GitHub page if it’s open-source. It’s great you have a domain name, but if your website is going to look the same as every other SaaS product designed by Claude it’s really hard to look past that and look at the novelty or benefits of the product.
I've built/am building something similar, but I spent the first half of my tech career as a UI/UX designer before becoming a software engineer and I'd _like_ to think it shows, but there is something about designing-in-code with agents that leads to homogenous outputs if you don't spend equal time on visual design as on the technical parts.
Looks great. I can tell you put a lot of time and energy into making it look good.
I think a lot of the problems with the homogenous outputs of front-end design wouldn't be such a problem if the models naturally make their designs so much simpler, but they are LLM's so they are always going to be overly verbose.
I was curious so I had asked my agent to redesign and recreate your front page for comparison and it gave me this: https://ouijit-redesign.vercel.app
These pages do look good. But they all just look the same. And I'm getting bored of them.
I open such a page and I immediately know it was Claude that produced it (probably end-to-end). Not that there's anything wrong with that, but it lacks soul… and that makes me kind of sad.
Tangential question for Claude Code subscribers, mid June `claude -p` will move to api pricing (with some "SDK credits" before it kicks in), so headless usage will become 20-30 times more expensive, and all these high level orchestrator tools/workflows depend on it. What the next move for you? How does the OpenAI subscriptions compare? Similar limitations?
The problem with this is that agents need to have good taste (code, design and UX) hammered into them with a crowbar and eben then writing manual CSS is something I find myself doing
Yes, this is like, the best thing ever .. I've generally been doing this, albeit with command-line Jira and a "my workflow is my prompt" philosophy, resulting in a fleet of little kanbans .. and my agents are really, really doing well. They never sleep, eat, etc.
But .. you know something cute? AI makes using Jira fun, again.
I feel 30 minutes of planning and 30 minutes of implementation in my solo side project's repo is too big to review. At minute 5, I may ask the AI to redo stuff even as its spitting out code.
reply