IT/Software career thread: Invert binary trees for dollars.

I've been using Claude to review PRs as a first line of defense. Using it inside VScode where it has context to the same codebase works really well. It can get pretty confused if your branch and the PR branch are way out of step.

I could easily see it getting to the point where AI writes the code, another agent review it, and another agent merges it with no human interaction.

Cad said:

Do you guys do that where you may use AI to completely write something or just assist and then put it into a fresh conversation and it finds new errors?

Click to expand...

That is fairly normal, and I'll do the same when I want to refresh the context window.

Think of the model as dividing its attention between every single thing you've talked about and shared up to that point. Every time you submit a new prompt, the model is generating its output by processing not just that prompt, but also the full history of your conversation--including all your prior prompts and the model's responses to generate each section, any back and forth you had editing or revising, and especially all those other case/fact docs you've exposed the model to. This basically dilutes the model's attention on just the brief text.

When you open a new window and say "Read this doc and provide critical feedback", the full weight of the model's output is generated solely against that prompt + the brief text, which means you get a much sharper response.

Cad said:

Do you guys do that where you may use AI to completely write something or just assist and then put it into a fresh conversation and it finds new errors?

Click to expand...

In addition to the context rot, I find that if it hits too much of a snag on something, it can get stuck in rabbit holes or decides that one path is the correct one and keeps coming back to it. A new session basically gets you a reroll on it's decisions. I'm only using it on relative hobby projects so far so so ymmv.

Deathwing said:

How does that(within VSCode) work if you want to containerize Claude to have some modicum of security?

BTW, the permissions you have to give Claude are fucking scary. Any other application, noping the fuck out for just 1 of those, let alone all 5.

Click to expand...

You can run it in a container or vm. A bit of extra hassle, but I don't want to completely turn it loose unless it's on a sacrificial system, and even then, careful about what it has access to on your network.

Boys, I had my second interview for a mini job (UiPath) yesterday. 15 hours per week, low pay, but despite that (sort of due to that) I really, really want this. I can get by just fine, and a low stress job sounds amazing. Never tried to be nervous for something like this before. I am about 10x overqualified, but it's all about culture fit and I wasn't the only one for second interview. I hope the rest are assholes.

Team up two and two and cross your wieners for me.

Deathwing said:

How does that(within VSCode) work if you want to containerize Claude to have some modicum of security?

BTW, the permissions you have to give Claude are fucking scary. Any other application, noping the fuck out for just 1 of those, let alone all 5.

Click to expand...

Claude is handled by our overlords here, but I work at a megacorp.

Also I agree on permissions. I think we very easily hand over the reins without thinking through the consequences sometimes.

I've been running the agents in VMs - good success with Multipass for headless Linux slices so far. I don't mount or share volumes between the host machine and the slices. You can setup a baseline Multipass dev enviroment, snapshot it, and then if you and/or the agent do something bad, you can just restore from snapshot. I figured this was worth the effort because I am still learning all these systems but I really wanted to let the agent be as autonomous as possible. The more autonomy they have, the more permissions they have, and the more of a chance something gets fucked up.

For repo access I've done it two ways. First way I just used Multipass file I/O to copy the repo over to the agent VM. I ask it to do some activity and it performs the edits. Works fine, but then you are the workhorse on repo management becaus you have to copy back the files, etc.

Second way is much easier but more setup. I created a Github account for the agent and then added that agent account as a collaborator to the repo I want it to work on. Make sure the repo has main branch protection setup. Then in the agent VM you just "gh auth login" like normal into that agent VM's account. Now you can tell it "Review code on repo XYZ" and it will clone the repo, etc. If you ask it to do edits, it will edit the code, you can review, and when you are happy you can say "Create a PR" and it will do it all.

With the second way, I have multiple agents setup with their own Github accounts. The farthest I have gotten right now is a code writer agent who makes the code edits and then a code review agent who only reviews PRs. Pretty funny to see a PR with two robots talking to each other.

Last thing - Visual Studio Code has an extension that lets you SSH into the VM so you can look at the code and more importantly the git diff to see what changes it made before it goes to a PR.

Deathwing said:

I'm being generally positive because I'm familiar with the code it's analyzing, I just spent hours/days working with the code. Thus, it's easy for me to understand its findings. I don't understand how people are using this to write feature-size code and have that same amount of familiarity with the results.

Click to expand...

Does it work? Yes? Who cares.

moonarchia said:

This is also exactly why there is money to be made unfucking what the AI fucks.

Click to expand...

I am betting that your typical human coders will 'fuck' things up just as much or more. The difference is that when AI does it, we can iterate and fix it stupidly fast and at almost zero marginal cost. Humans? Good luck getting them to even admit the bug exists, let alone ship the fix before the next sprint ends.

I think he was talking about actual indians

Haus said:

My inner cynic has to chime in on this one...

Training a bunch of fresh out of college kids means getting them up to speed to handle the essentials, plus use AI to accelerate their ramp up. Then what do you have?

A bunch of kids who don't know what market value for good coders is and a bunch of seasoned developers, allowing you to them stop giving raises, or even treating the experienced folks well. As they leave, the kids step in.

Don't ask me how I have such knowledge of this

On the more positive/optimistic side. Some places are realizing the only way you even develop someone with both deep talent and knowledge of your organization is to grow them "from a seedling".

Click to expand...

The cynical take assumes that the current pool of human developers, even seniors and architects, are consistently making high-quality decisions. From what I’ve seen in our own codebase, that’s often not the case.

We have multiple examples of questionable design choices and technical debt introduced by experienced engineers, frequently based on outdated patterns or "how we’ve always done it." These aren’t rare mistakes, they’re real issues in our application (a good example is a senior architect insisted on not using the async/await pattern for our .NET API because it made 'debugging difficult' and now our API has severe performance issues under high-load.)

This is why I see more potential in AI coding agents than some critics do. They’re trained on enormous datasets of modern code, widely accepted best practices, and current frameworks. They don’t carry the same personal baggage or legacy thinking that even seasoned developers sometimes do.

They’re obviously not perfect. They can generate slop and lack deep understanding of a specific domain and system. But neither are humans. At least with AI you get consistency and exposure to up-to-date patterns by default.

I’m not saying juniors + AI will automatically outperform great seniors. I’m saying we shouldn’t romanticize the existing senior talent as flawless, because they aren’t.

Dune server