Head of Engineering, 18 years exp. AI agents in a 40-repo production codebase + side projects. The cost, the hit rate, and the parts that break
🇧🇷🇳🇱🇪🇸🇺🇸
@naveen_saradhi_ what about interacting with AI to design the schema the you want, other than doing manually? ofc you can do it manually, but feels more like a hobby other than strictly necessary.
@tchaguitos Boa! Tambem tenho feito isso, o unico problema é quando tem muito PR pra validar né? Ou quando agente volta com alguma pergunta mid-flight 😅
Def true, and I would add the boring half of why it works.
These past weeks I had Claude map out the UI test scenarios on a project and run them. It comes back with the scenario that failed instead of me guessing from a stack trace. That is what actually changed for me. Not that the agent writes better code, but that I find out faster when it did not.
The prod horror stories almost always have that step missing. I would love to know what Stripe has sitting underneath that 1,300 number.
Uso Claude Code todo dia e mantenho as specs como cards, porque pedido vago faz
ele ler o repositório inteiro antes de começar a trabalhar. Preparo o card com outros modelos e deixo o Fable para implementação. Isso tem me ajudado a nao exceder o limite. Alem de ter algumas skills com contextos especificos de partes do sistema.
Same here, I was using WisprFlow for a while. Its nice that it can integrate with everything. But I also just feel like typing still much better.
I think a nice use case would be, if you are in the middle of a task, and the agent asks you in voice mode what you want to do about a mail, or a notification you just received. Then you can easily reply back, without completely shifting your attention to where it came from.
Like "Team Member A sent you a message about X",
"Ok, tell you him he can find it at Y"
I don't believe too much in using for every day's tasks.
@kinder_grinder There was nothing suspicious in the logs?
Couldn't it be a dependency issue? The other day I had an issue at the restoring packages phase. I had a .Net application, and some nuget packages sources were unstable. Nothing I could do.
These past weeks I've been instructing agents to map-out all possible test cases of projects I'm working on. My agent performs all UI test cases, in case something breaks, it creates a task in the board with all details of the problem, and it because it also knows the code base it also knows how to pin point the problem. Other than creating visual evidence of the issue.
Instructing the agent to create output files to tell you why it broke in the testing phase, in a technical/or non technical way it's also possible.
There are still teams out there using white boards with yellow stickies to manage their sprints and iterations.
Have you ever gave it a try using a board like Notion, Jira, or Azure DevOps, and asking your agent to keep updating it as you go?
Nice plus of this approach is that you can also instruct your agent to enrich your ideas and update it for you.
Pretty easy setup.
@eddiejaoude Google Stitch, Azure CLI to fast prosivion resources in the cloud, Azure Board to manage the work (even though the UI is not great, is free up to 5 users), Claude, and VSCode. Stripe cli if any payment gateway is needed.
I haven't tried @ElevenLabs yet. I actually did something very simple the other day, also with Claude. I recorded myself using Photo Booth, asked Claude to transcript what I said, it used mac's internal transcript tooling for that, then I just asked to add to the video I had recorded. The downside might be Claude might consume more tokens versus just using a specific tool for this.