Oscar Romero

Ex-CTO, exit. 7 yrs taking AI from demo to production. Fractional CTO + freelance: LLM, agents, Python. Obsessed with the AI context.

[Oscar Romero @ORomL]

@karpathy Soft. Engineer and Farmer, also. Tried to drop a prompt to build a small ERP for my farms to track production.

BUT, with good software practices that i know as a senior dev/cto.

Result? Barely working code that required multiple reviws and iterations to move to prod.

[Oscar Romero @ORomL]

Give a model complete, relevant information and its performance jumps dramatically.

For many real-world tasks, the bottleneck is no longer reasoning.

The next AI race may not be about building smarter models.

It may be about building better context.

[Oscar Romero @ORomL]

Everyone is talking about the US cutting off access to Fable/Mythos.

I had the chance to test it myself, and honestly, with proper context management, the gap vs Opus is much smaller than most people think.

That led me to question while developing https://t.co/rbJkdRClHQ...

[Oscar Romero @ORomL]

I'm currently building https://t.co/rbJkdRClHQ, a system that helps AI agents access information scattered across multiple sources.

In my tests, both Fable and Opus were able to retrieve all project-related documents in a single call and successfully complete an entire sprint.

[Oscar Romero @ORomL]

@CooperTechHub "Better Context Management" here is where Im working on with https://t.co/rbJkdRClHQ the main idea is centralize context to allow agents by just 1-2 tool calls extract all relevant information from multiple sources.

[Oscar Romero @ORomL]

@chaosengineerr Actually building https://t.co/rbJkdRClHQ that is a context layer for agents.

[Oscar Romero @ORomL]

@jam_thecreator @clarefinds Thx @jam_thecreator I will check it out.

[Oscar Romero @ORomL]

@nikks_techie Haaay ! 👋

[Oscar Romero @ORomL]

Where I found alot of difference is using my project https://t.co/rbJkdRClHQ that reduces the tool calling from 5 to 1 for each task reducing the token usage and speedup development.

[Oscar Romero @ORomL]

Today im reading alot about the USA policies about the Fable 5 model. To be honest with a good context management and well defined workflows I didnt find alot of breaking differences between Opus, Gpt 5.5. or Fable.

[Oscar Romero @ORomL]

@_oliveiradanilo This happened to a customer that Im working with. His problem is app side that was generating and resuing the same cache all time that forces Google to keep the cache live without stop, applying costs for it. The solution in his case whas a better cache policy on the app side.

[Oscar Romero @ORomL]

@compileandpush Depends on the context and the problem itsef.

[Oscar Romero @ORomL]

Spent the last few days building a new project: https://t.co/rbJkdRClHQ
The nice part: my coding agents use the project itself to pull context about the project. Very ouroboros.
The not-so-fun part 👇
#buildinpublic #AgenticCoding

[Oscar Romero @ORomL]

Moral of the story: time to add more conditions and gates to my skills.
Coding agents are great. But all that glitters is not gold — sometimes it's just a very confident loop digging a hole.
#buildinpublic #AI #devtools

[Oscar Romero @ORomL]

Yesterday CI broke. The infrastructure...
My agents didn't care. They went full "make it green" mode and started shipping useless fixes.
By the time I noticed, they'd broken working code, blown up their context windows, and literally forgotten how to use their own skills. 🫠

[Oscar Romero @ORomL]

@claudeai @OpenAI @GeminiApp This increases the operation cost of them. We need to rethink tool-rl and agentic-rl methods to reason before tool use.

[Oscar Romero @ORomL]

@claudeai is happening also with GPT 5.5 from @OpenAI and Gemini 3.5 from @GeminiApp the new "agentic models" are starting have a compulsive behavior to use all available tools before reasoning.

[Oscar Romero @ORomL]

@thegautamkamath Is normal... With this picture you said that your wife is a transformer! 🤣