Evandro B Filho

Software Technical Lead @MercadoLivre. Building AI agents and developer tools. 10+ years with Java, Kotlin, Scala, Kafka and high-throughput systems.

[Evandro B Filho @evandrocabf]

@GoogleCloudTech Wouldn’t the 32,768-token cutoff move a lot with cache TTL and model pricing, instead of being a fixed rule?

[Evandro B Filho @evandrocabf]

LLMs made unfamiliar syntax cheap. Runtime quirks did not get a discount.

Would you choose the same language after the first weird failure?

[Evandro B Filho @evandrocabf]

Three weeks between model names. Somewhere, a dropdown just became legacy software.

Build around behavior you can test, not branding you have to chase.

https://t.co/fbHWNDMSRF

[Google @Google]

Introducing Gemini 3.8, our best reasoning & coding model yet.

By leveraging long-running agentic loops, we’re building on the momentum of 3.7 Flash from just three weeks ago to release two new 3.8 variants:

Meet Gemini 3.8 Flash and Gemini 3.8 Flash Cyber 🧵

[Evandro B Filho @evandrocabf]

@arena @GoogleDeepMind The #7 Text vs #18 WebDev gap makes me curious where it loses points: planning, tool use, or recovering after a bad edit?

[Evandro B Filho @evandrocabf]

When you know the symbol, embeddings are fancy Ctrl+F. When you only know the intent, grep is a detective with no clues.

Good code search should stop making builders choose.

[Evandro B Filho @evandrocabf]

A rewrite gets a clean architecture diagram. A migration gets shared state, old lifecycle rules, and users who expect Tuesday to work like Monday.

Guess which one teaches you more about the system.

[Evandro B Filho @evandrocabf]

@fernandorojo Will model choice stay pinned per project, so a build doesn’t quietly change style when the default changes?

[Evandro B Filho @evandrocabf]

A coding session does not resume because fresh tokens arrived. It resumes when the plan, last known state, and next action survive the pause.

Anything less is amnesia with a retry button.

[Evandro B Filho @evandrocabf]

More MCP tools under one giant permission scope turn capability into blast radius. Nice toolbox, terrible key management.

How little access should an agent need to finish the task?

[Evandro B Filho @evandrocabf]

Routing across AI providers looks like optimization until one fails. Then it looks like uptime. Boring fallbacks keep winning.

[Evandro B Filho @evandrocabf]

The model can write the feature before the team agrees what the feature should do. A vague ticket is now the first bug in the repo.

[Evandro B Filho @evandrocabf]

With 200+ open models running at the metro edge, location becomes a runtime choice.

Builders need routing and observability by region, not just a model dropdown.

https://t.co/Vj7LceFWNX

[Evandro B Filho @evandrocabf]

@marclou Does the MCP load those 60+ tools by group, or are all schemas entering the agent context at once?

[Evandro B Filho @evandrocabf]

Price per token is a cute metric until a small task turns into a novel.

What would model leaderboards look like if they ranked cost per accepted change instead?

[Evandro B Filho @evandrocabf]

AI coding demos end at the first answer. Building software starts there.

The useful agent is not the one that looks smartest. It is the one that leaves a diff, tests, and proof you can trust.

[Evandro B Filho @evandrocabf]

@thsottiaux Am I using codex remote properly? Or did I get it wrong?

[Evandro B Filho @evandrocabf]

@FelipeFr1702 Use Sol High, and Terra, stop using Max hahah.

[Evandro B Filho @evandrocabf]

@FelipeFr1702 Hahaha, I've abused of it a bit too, but I'm still at 52%.

[Evandro B Filho @evandrocabf]

@thsottiaux When we will have GPT 5.7 betelgeuse?

[Evandro B Filho @evandrocabf]

@omarsar0 Does the preference model reserve some GPU budget for low-ranked ideas, so its taste doesn't become too conservative?