Blog | Juanchi.dev | Juanchi.dev

Tag: agentes-ia

27 articles

Real guardrails for autonomous agents after one almost destroyed my infrastructure

After an autonomous agent nearly wiped my production database, I built a real guardrails layer. Here are the controls, the code, and the logs that saved my skin.

6 min182

ExperimentsTypeScriptdevops

Agents That Create Accounts, Buy Domains, and Deploy on Their Own: I Tested It Against My Real Stack — Here's What Broke (and What Worked)

The viral HN demo shows Cloudflare agents running the full infra cycle with zero human intervention. I replicated it against my Railway stack and documented exactly what the agent executed on its own, where I had to step in, and what permissions it asked for that it absolutely shouldn't have. Real a

9 min188

ExperimentsTypeScriptclaude code

DeepClaude: I Combined Claude Code with DeepSeek V4 Pro in My Agent Loop and the Numbers Threw Me Off

I took the DeepClaude repo (467 points on HN) and dropped it into my real production loop. The combination isn't simply "better than either alone" — there's a specific task regime where DeepSeek V4 Pro destroys and Claude fails, and vice versa. Here are my actual numbers.

8 min235

ExperimentsNext.jsTypeScript

Specsmaxxing: I Wrote YAML Specs for My AI Agents — Here's What Changed (and What Didn't)

Specsmaxxing promises to cure "AI psychosis" with YAML specs for agents. I applied it to my real workflow with Claude Code and found the trap nobody mentions: the quality problem doesn't disappear — it moves into the YAML.

8 min247

ExperimentsTypeScriptclaude code

I tried to reproduce the OpenClaw case in Claude Code: my result contradicts the viral post

The HN thread claimed Claude Code blocked or redirected billing when OpenClaw appeared in Git history. I built a public repo, a reproducible harness, and ran the matrix. On Claude Code 2.1.126, I did not reproduce the block.

5 min195

OpinionTypeScriptclaude code

Who owns the code Claude Code wrote? I ran git blame on a real project and the result is uncomfortable

I ran git blame on a project where I used Claude Code heavily. 61% of the lines aren't mine. That's not a legal problem yet — it's an accountability problem for when something blows up in production and nobody knows whose signature is on it.

8 min132

ExperimentsTypeScriptrailway

GPT-5.5 in the API: I ran it against my real production cases and the numbers don't justify the upgrade yet

I ran GPT-5.5 against my actual production prompts and compared it to GPT-4o on latency, cost, and output quality. The marketing leap doesn't match the leap in my metrics. Here are the numbers.

8 min149

ExperimentsTypeScriptclaude code

I Almost Cancelled Claude: I Ran My Own Benchmarks Before Pulling the Trigger

874 points on HN for 'I cancelled Claude'. Before joining the chorus, I ran my own regression cases against real Claude Code logs. The degradation is real — just not where everyone's complaining about it.

8 min137

ExperimentsTypeScriptproduccion

LLMs generating security reports: I ran the same prompt on reproducible example code

HN reported that the Linux kernel is receiving removals based on LLM-generated security reports. I took the same pattern and ran it against my own production code. What I found made me uncomfortable — but not for the reasons I expected.

9 min121