Diego Prieri

Lead Frontend & AI Engineer. Building agent workflows on Claude Code and measuring if they work. 87% lower LLM cost, 86% first-try typecheck. Buenos Aires

[Diego Prieri @Diegofush]

Hi X 👋 I’m Diego, a frontend engineer in Buenos Aires. I lead a ~120k-line Next.js app where AI agents write most of the code, and I build open-source tools to measure whether that actually works.

If you’re building with Claude Code, MCP or coding agents, say hi.

[Diego Prieri @Diegofush]

@m_sakibnaz Nice, an iPhone grip is a cool niche! Agents on MCUs is the harder version of my problem: on the web I just run the tests, but on hardware, how does the agent know the firmware works? Simulator, a real board in the loop, or both?

[Diego Prieri @Diegofush]

Building with AI agents? Let’s talk 👋

I lead a Next.js app where AI agents write most of the code, and I build tools to measure if it works.

Looking to meet:

💻 Frontend devs
🤖 AI engineers
🛠️ Claude Code & MCP builders
🚀 B2B SaaS teams
Drop a “Hey” or share what you’re building!

[Diego Prieri @Diegofush]

@LiveFrom2100 Agreed, that part is unsolved: a test can’t prove a human understood the diff. My numbers are here, happy to go deeper: https://t.co/QqJ9TNffX9

[Diego Prieri @Diegofush]

Hi X 👋 I’m Diego, a frontend engineer in Buenos Aires. I lead a ~120k-line Next.js app where AI agents write most of the code, and I build open-source tools to measure whether that actually works.

If you’re building with Claude Code, MCP or coding agents, say hi.

[Diego Prieri @Diegofush]

@LiveFrom2100 Ha, a scary headline, but plausible. My bet: responsibility stays with whoever merged it. That’s why I want evidence at merge time: tests ran, someone understood the diff. Happy to share real numbers if you write that dispatch.

[Diego Prieri @Diegofush]

@tazqen Nice, diff and preview next to the session is where review should live. One thing I’d want there: what the agent ran before asking. Did tests pass, or did it just say done? Then the human decides with evidence, not just the diff.

[Diego Prieri @Diegofush]

@RealUgboma Exactly. The easy part: agent commits already carry a Co-Authored-By trailer, and session timestamps match the commit time. The hard part is “fix”: a later commit touching the same lines, or only ones that say fix/revert? Still deciding.

[Diego Prieri @Diegofush]

@KelvenSeveroF Ran it. Tracked .ts files only, the typecheck gap flips: edit tools 80.7% (n=83) vs Bash 85.6% (n=181). So it wasn’t a real advantage, mostly scratch files. Caveat: git only knows today’s files, so deleted worktrees drop out. Posting either way, as promised.

[Diego Prieri @Diegofush]

@NbSoloman Hey, welcome! Frontend lead from Argentina here, working on AI agents in production code and measuring whether their output holds up. What are you building in Hong Kong? Happy to connect 🤝

[Diego Prieri @Diegofush]

@tazqen Thanks, glad it resonates! Visible review is underrated: when an agent writes most of the code, the review is where the real decisions happen. What does “visible” mean in Tazqen? Review in the PR, or something built into the agent loop?

[Diego Prieri @Diegofush]

@LiveFrom2100 Thanks, happy to connect! Long-term consequences are the part nobody measures yet. With code, I wonder what happens to a codebase after years of mostly agent-written commits. Is that kind of question in scope for your newsroom?

[Diego Prieri @Diegofush]

@getnachonacho Thanks! Nothing exotic: tsc, Jest with 100% coverage, and Playwright for E2E. Pre-commit hooks run them so the agent can’t skip them, and a separate review agent checks each diff. Then claude-code-metrics reads the transcripts to see how often the first run passes.

[Diego Prieri @Diegofush]

@Davidariasfin Comparing LLMs on the same repos is exactly the right setup. With LLMs the look-ahead problem is worse: the model may have seen the backtest period in training. Do you only test on dates after each model’s cutoff? Would love to see the runs database.

[Diego Prieri @Diegofush]

Las escrituras de caché con TTL de 1 hora superan a las de 5 minutos 2,6 a 1 (84.7M contra 32.5M tokens), y aun así conviene.

[Diego Prieri @Diegofush]

@KelvenSeveroF Agreed, that’s the right control. Shell writes include scratch files and generated code that never reach the repo, which could explain the gap. I’ll rerun with tracked .ts files only and post what changes, even if the gap disappears.

[Diego Prieri @Diegofush]

@Orvi_dev Let’s connect! Safety in orchestration is the part I find least solved. For me it means gates the agent can’t skip, like tests before commit. What does safety look like at noburn? Limits on actions, or review?

[Diego Prieri @Diegofush]

@valentinoswe Genial! Named steps that only advance on green checks is a clean model. For progress inside a step, what I count is check runs until green: first try is 1, a long fix loop shows up right away. Could be a cheap signal for vstack.

[Diego Prieri @Diegofush]

@Davidariasfin Hey, let’s keep in touch! Financial markets is where verification gets real: a wrong answer there costs money, not just a failed test. How do you check an agent’s decision before it acts? Backtests, human approval, both?

[Diego Prieri @Diegofush]

@peteryueai Hi Peter, happy to connect! AI marketing employees sounds fun to build. How do you check their output before it goes out? With code I have tests. With copy, “good enough” seems much harder to measure.

[Diego Prieri @Diegofush]

@m_sakibnaz Thanks for the follow, followed back! Hardware is a tough one to go solo on. What are you building? Curious whether you use AI agents for the firmware side too, or mostly for the business side.

[Diego Prieri @Diegofush]

@chngsng354982 Nice, let’s connect! I’m mostly on agents in production code and measuring if their output holds up. What are you shipping? Always curious what solo builders automate first.

[Diego Prieri @Diegofush]

@MichaelZima Thanks! The feedback loop is the part I care most about. What are you building? Happy to compare notes if it’s agents or verification.