Comment by defatigable | Hacker News Reader

Comment by defatigable | Hacker News Reader

Parent Full thread

I use Augment with Claud Opus 4.5 every day at my job. I barely ever write code by hand anymore. I don't blindly accept the code that it writes, I iterate with it. We review code at my work. I have absolutely found a lot of benefit from my tools.

I've implemented several medium-scale projects that I anticipate would have taken 1-2 weeks manually, and took a day or so using agentic tools.

A few very concrete advantages I've found:

* I can spin up several agents in parallel and cycle between them. Reviewing the output of one while the others crank away.

* It's greatly improved my ability in languages I'm not expert in. For example, I wrote a Chrome extension which I've maintained for a decade or so. I'm quite weak in Javascript. I pointed Antigravity at it and gave it a very open-ended prompt (basically, "improve this extension") and in about five minutes in vastly improved the quality of the extension (better UI, performance, removed dependencies). The improvements may have been easy for someone expert in JS, but I'm not.

Here's the approach I follow that works pretty well:

1. Tell the agent your spec, as clearly as possible. Tell the agent to analyze the code and make a plan based on your spec. Tell the agent to not make any changes without consulting you.

2. Iterate on the plan with the agent until you think it's a good idea.

3. Have the agent implement your plan step by step. Tell the agent to pause and get your input between each step.

4. Between each step, look at what the agent did and tell it to make any corrections or modifications to the plan you notice. (I find that it helps to remind them what the overall plan is because sometimes they forget...).

5. Once the code is completed (or even between each step), I like to run a code-cleanup subagent that maintains the logic but improves style (factors out magic constants, helper functions, etc.)

This works quite well for me. Since these are text-based interfaces, I find that clarity of prose makes a big difference. Being very careful and explicit about the spec you provide to the agent is crucial.

This. I use it for coding in a Rails app when I'm not a Ruby expert. I can read the code, but writing it is painful, and so having the LLM write the code is beneficial. It's definitely faster than if I was writing the code, and probably produces better code than I would write.

I've been a professional software developer for >30 years, and this is the biggest revolution I've seen in the industry. It is going to change everything we do. There will be winners and losers, and we will make a lot of mistakes, as usual, but I'm optimistic about the outcome.

Agreed. In the domains where I'm an expert, it's a nice productivity boost. In the domains where I'm not, it's transformative.

As a complete aside from the question of productivity, these coding tools have reawakened a love of programming in me. I've been coding for long enough that the nitty gritty of everyday programming just feels like a slog - decrypting compiler errors, fixing type checking issues, factoring out helper functions, whatever. With these tools, I get to think about code at a much higher level. I create designs and high level ideas and the AI does all the annoying detail work.

I'm sure there are other people for whom those tasks feel like an interesting and satisfying puzzle, but for me it's been very liberating to escape from them.

> In the domains where I'm an expert, it's a nice productivity boost. In the domains where I'm not, it's transformative.

Is it possible that the code you are writing isn't good, but you don't know it because you're not an expert?

No, I'm quite confident that I'm very strong in these languages. Certainly not world-class but I write very good code and I know well-written code when I see it.

If you'd like some evidence, I literally just flipped a feature flag to change how we use queues to orchestrate workflows. The bulk of this new feature was introduced in a 1300-line PR, touching at least four different services, written in Golang and Python. It was very much AI agent driven using the flow I described. Enabling the feature worked the first time without a hiccup.

(To forestall the inevitable quibble, I am aware that very large PRs are against best practice and it's preferable to use smaller, stacked PRs. In this case for clarity purposes and atomicity of rollbacks I judged it preferable to use a single large PR.)

> I've implemented several medium-scale projects that I anticipate would have taken 1-2 weeks manually

A 1-week project is a medium-scale project?! That's tiny, dude. A medium project for me is like 3 months of 12h days.

You are welcome to use whatever definition of "small/medium/large" you like. Like you, 1-2 weeks is also far from the largest project I've worked on. I don't think that's particularly relevant to the point of my post.

The point that I'm trying to emphasize is that I've had success with it on projects of some scale, where you are implementing (e.g.) multiple related PRs in different services. I'm not just using it on very tightly scoped tasks like "implement this function".

I mean, if it's working for you, great.

The observation I was trying to make is that at the scope of one week, there's very little you actually get done, and it's likely mostly mechanical work. Given that, I suppose I'm unsurprised LLMs are proving useful. Seems like that's the type of thing they're excelling at.

That's not my experience. I agree that a project of any real size takes quite a bit longer than a week. But it's composed of lots of, well, week or two long subprojects. And if the AI coding tool is condensing week long projects into a day, that's a huge benefit.

Concretely speaking (well as concretely as I feel like being without piercing pseudonymity), at my last job I worked on a multi year rewrite of one of our core services. Within that rewrite were ton of much smaller projects that were a few weeks to a month long - refactor this algorithm, improve the load balancing, add a new sharding strategy, etc. An AI tool would definitely not have sped up the whole process. It's not going to, say, speed up figuring out and handling intra-team dependencies or figuring out product design. But speeding up those smaller coding subprojects would have been a huge benefit.

I'm not making any strong claims in my post. I don't have the experience of AI projects allowing me to one shot large projects. But OP asked if anyone has concrete experience with AI coding tools speeding up development, and the answer is yes, I do.

Well a medium project for me takes 3 years, so obviously I am the best out of everyone /s

1. And 2. I.e. creating a spec which is the source of truth (or spec driven development) is key to getting anything production grade from our experience.

Yes. This was the key thing I learned that let me set the agents loose on larger tasks. Before I started iterating on specs with them, I mostly had them doing very small scale, refactor-this-function style tasks.

The other advice I've read that I haven't yet internalized as much is to use an "adversarial" approach with the LLMs: i.e. give them a rigid framework that they have to code against. So, e.g., generate tests that the code has to work against, or sample output that the code has to perfectly match. My agents do write tests as part of their work, and I use them to verify correctness, but I haven't updated my flow to emphasize that the agents should start with those, and iterate on them before working on the main implementation.

I wouldn't consider the proposed workflow agentic. When you review each step, give feedback after each step, it's simply development with LLMs.

Interesting. What would make the workflow "agentic" in your mind? The AI implementing the task fully autonomously, never getting any human feedback?

To me "agentic" in this context essentially that the LLM has the ability to operate autonomously, so execute tools on my behalf, etc. So for example my coding agents will often run unit tests, run code generation tools, etc. I've even used my agents to fix issues with git pre-commit hooks, in which case they've operated in a loop, repeatedly trying to check in code and fixing errors they see in the output.

So in that sense they are theoretically capable of one-shot implementing any task I set them to, their quality is just not good enough yet to trust them to. But maybe you mean something different?

IMHO, agentic workflow is the autonomous execution of a detailed plan. Back-and-forth between LLM and developer is fine in the planning stage. Then, the agent is supposed to overcome any difficulties or devise solutions to unplanned situations. Otherwise, Cursor had been able to develop in a tight loop of writing and running tests, followed by fixing bugs, before “agentic” became a buzzword.

Perhaps “agentic” initially referred to this simple loop, but the milestone was achieved so quickly that the meaning shifted. Regardless, I could be wrong.

Yeah, I have no idea what the consensus definition of the term is, and I suppose I can't say for sure what OP meant. I haven't used Cursor. My understanding was that it exercises IDE functions but does not execute arbitrary shell commands, maybe I'm wrong. I've specifically had good experiences with the tools being able to run arbitrary commands (like the git debugging example I mentioned).

In my experience reading discussions like this, people seem to be saying that they don't believe that Claude Code and similar tools provide much of a productivity boost on relatively open ended domains (i.e. the AI is driving the writing of the code, not just assisting you in writing your own code faster). And that's certainly not my experience.

I agree with you that success with the initial milestone ("agent operates in a self-contained loop and can execute arbitrary commands") was achieved pretty quickly. But in my experience a lot of people don't believe this. :-)

Same, Opus 4.5 is nothing short of amazing. I’m really shocked to see so many posts claiming it doesn’t work.

We write whole full scale Rust SaaS apps with few regressions.

I do novel machine learning research in about a 1/10 of the time it would have taken me.

A big thing is telling it to excessively log so it can see the execution

Great advice.

> Tell the agent your spec, as clearly as possible.

I have recently added a step before that when beginning a project with Claude Code: invoke the AskUserQuestionTool and have it ask me questions about what I want to do and what approaches I prefer. It helps to clarify my thinking, and the specs it then produces are much better than if I had written them myself.

I should note, though, that I am a pure vibe coder. I don't understand any programming language well enough to identify problems in code by looking at it. When I want to check whether working code produced by Claude might still contain bugs, I have Gemini and Codex check it as well. They always find problems, which I then ask Claude to fix.

None of what I produce this way is mission-critical or for commercial use. My current hobby project, still in progress, is a Japanese-English dictionary:

https://github.com/tkgally/je-dict-1

https://www.tkgje.jp/

Great idea! That's actually the very next improvement I was planning on making to my coding flow: building a sub agent that is purely designed to study the codebase and create a structured implementation plan. Every large project I work on has the same basic initial steps (study the codebase, discuss the plan with me, etc) so it makes sense to formalize this in an agent I specialize for the purpose.

Is it just me, or does every post starting with "Great Idea!" or "Great point!" or "You're so right!" or similar just sound like an LLM is posting?

Or is this a new human linguistic tic that is being caused by prolonged LLM usage?

Or is it just me?

:-) I feel you. Perhaps I should have ended my post with "Would you like me to construct a good prompt for your planning agent?" to really drive us into the uncanny valley?

(My writing style is very dry and to the point, you may have noticed. I looked at my post and thought, "Huh, I should try and emotionally engage with this poster, we seem like we're having a shared experience." And so I figured, heck, I'll throw in an enthusiastic interjection. When I was in college, my friends told me I had "bonsai emotions" and I suppose that still comes through in my writing style...)

Excellent reply :) And yes, maybe that's it, that the LLM emotion feels forced so any forced emotion now feels like an LLM wrote it.

Reply on news.ycombinator.com