esteban 🇺🇦

エステバン – | Software Engineer at @lancedb. HBase Committer, ex-{Datahub, @redpandadata, @Cloudera, @SismologicoMX,@cires_ac,@GobCDMX}. Swim dad. All views mine.

[esteban 🇺🇦 @esteban]

I’m sure nobody thought about the terrible design choice as X the logo vs x the the app name. I’m uninstalling this, find me in @threads or @mastodon

[Lucas Maes @lucasmaes_]

Most work still relies on MP4 videos ... creating a major bottleneck that kills GPU utilization and training throughput.

Not anymore.

We teamed up with the incredible @lancedb team to build blazing-fast data loading with native @huggingface bucket streaming support! ⚡

[Lucas Maes @lucasmaes_]

Would you like to join the research effort on JEPA and World Models easily?

After a full year of hard work, we’re excited to finally release stable-worldmodel:

an open-source, scalable platform built to accelerate JEPA & World Model research!

📄: https://t.co/gnxGvens5A

[esteban 🇺🇦 @esteban]

I’ve been saying this for months: we are all running into the same patterns and is a shame that our progress is limited to our own repos or laptops. I’ve been doing this for the last 2 months and I’m glad to see Anthropic making this a proper feature.

[ClaudeDevs @ClaudeDevs]

New in Claude Code (research preview): dynamic workflows.

Claude writes an orchestration script on the fly, then spins up a large fleet of coordinated subagents in parallel to take on your most complex tasks.

Use the word "workflow" in a prompt to get started.

[esteban 🇺🇦 @esteban]

@bcherny Cool! this is what I’ve been doing for months and I’m glad is now a proper feature.

[esteban 🇺🇦 @esteban]

@bcherny So can i remove all my introspection rules now? I feel that asking Opus 4.7 to “don’t rush” “double check, when you say is done it must be done for real after running unit tests, etc” was a waste of tokens.

[esteban 🇺🇦 @esteban]

I feel that I should release my framework after this or at least part of it. The convergence of what we are discovering what models are capable to introspect about their own work is remarkable.

[Muratcan Koylan @koylanai]

Gradient descent for SKILL.md files sounds interesting, maybe a bit complex but it's becoming a real part of agent harness.

SkillOpt is one of the first papers to treat markdown skill files as trainable parameters and provides a proper optimization framework for them.

A few things I learned that you should consider too.

1. The validation gate is the only thing that matters in a self-editing loop.

Held-out set, strict improvement, ties rejected. End-to-end, their best skills land with 1 to 4 accepted edits total. If your "self-improving agent" is accepting most of what it proposes, you're shipping slop.

2. Bounded edits are better than full rewrites. 4 to 8 edits per step is the sweet spot.

Remove the budget and performance collapses. This is the textual analog of learning rate, and it transfers to any LLM-as-author loop. If you're using an agent to refactor your docs, your prompts, or your skills, cap the diff size.

3. Compactness wins. Median final skill: ~920 tokens.

Skills do not need to be long. They need to be high-signal. Most skill files I see are bloated because length feels like effort. It isn't.

4. The harness is becoming less important; the skill is becoming more important.

A Codex-trained skill ported into Claude Code hit +59.7 points on SpreadsheetBench. Procedural knowledge is more general than the runtime that
produced it.

5. Frozen model + trained context is the practical adaptation.

GPT-5.4-nano with a SkillOpt'd skill ≈ frontier behavior on procedural benchmarks. Cheaper, portable, inspectable, zero inference-time cost. This is
the answer to "how do we adapt a frontier model for our domain" for almost everyone who isn't training their own models.

6. Verification is the bottleneck.

Every gate in this paper depends on an auto-grader. That works for benchmarks. It fails for writing, design, and strategy, exactly the open-ended work we want to automate. Whoever builds the verifier for open-ended tasks owns the next stage.

There are also two leassons I learned while shipping v2.3.0 of my Context Engineering Agent Skills repo, measured across composer-2, claude-opus-4-7,
gpt-5.5, and gemini-3.1-pro via the @cursor_ai SDK:
- Description and body are two different surfaces. The router only sees the description. The agent sees the body once activated. They can quietly disagree, and only end-to-end task tests catch it.
- Aggregate accuracy is the wrong unit. When I rewrote three descriptions, the corpus average moved ~1pp. Individual skills moved 23–25pp. Per-skill effect size is where the action is.

Also, in Feb 2026 I shared a piece called Personal Brain OS arguing that the markdown file is a first-class substrate for agent state. SkillOpt is the optimizer-shaped version of that same argument: not "store memory in files" but "treat files as trainable parameters with proper optimization machinery around them." That's the move from static to measured.

The fast/slow split they describe already lives implicitly in the digital-brain-skill repo:
- voice-guide and tone-of-voice.md are slow-state (rarely touched)
- posts.jsonl and bookmarks.jsonl are fast-state

What SkillOpt adds that I didn't have is a protected section invariant, a structural guarantee that fast edits cannot overwrite slow lessons. Removing that mechanism cost them 22 points on SpreadsheetBench. Worth borrowing.

If you're building agents, SkillOpt: Executive Strategy for Self-Evolving Agent Skills is a good paper to read: https://t.co/ZS9SZXQ6Mv

[esteban 🇺🇦 @esteban]

Not really. I always wanted to write scientific software until I fell in love with databases and distributed systems.

[Kelindi @_kelindi]

every software engineer is just someone who wanted to make video games and quietly gave up on that dream

[Aaron Levie @levie]

CEOs are uniquely prone to AI psychosis because they’re sufficiently distant from the last mile of work that still has to happen to generate most value with AI.

So when they play with AI, they see the happy path results, often not considering the next 10 or 20 things that have to happen to get sustainable results from agents.

“Look I made this awesome product prototype”. Yes but you didn’t have to review the code before it went into production and fix a bunch of issues.

“Look I generated a contract”. Yes but you didn’t verify all the terms before it goes out to the counterparty and didn’t have to wire up all the past contracts to work with.

The best thing you can do as a CEO is to use AI a *ton* to figure out the real implications of agents in the enterprise, and come out the other side with an appreciation for both the upside and the real work that goes into them.

[esteban 🇺🇦 @esteban]

IYKYK

[Corey Quinn @QuinnyPig]

@AWSOpen Still waiting on merge...

https://t.co/IC1oG7qpA0

[esteban 🇺🇦 @esteban]

@ReikoSPCT you too!

[esteban 🇺🇦 @esteban]

@ReikoSPCT 1. yes
2. yes
3. Cedar Park, TX, 5/21/2026 20:23

[esteban 🇺🇦 @esteban]

@StringerViral @averytomascowx @HEB you have my permission.

[esteban 🇺🇦 @esteban]

@PowellGeorge you have my permission.

[esteban 🇺🇦 @esteban]

@atxwxgirl enjoy!

[esteban 🇺🇦 @esteban]

@averytomascowx @HEB parking lots never disappoint for this.

[esteban 🇺🇦 @esteban]

@QuinnyPig I need that prompt.

[Avery Tomasco @averytomascowx]

TIME-LAPSE: WOOSH

What a storm in Austin! Strong winds have knocked out the power to more than 15,000 Austin Energy customers. That's more than Tuesday's severe storm #atxwx #txwx

[Xuanwo @OnlyXuanwo]

. @lancedb 💕 @duckdb

[esteban 🇺🇦 @esteban]

@HEB @NWSSanAntonio

[esteban 🇺🇦 @esteban]

How is your fire rainbow doing Austin?

[esteban 🇺🇦 @esteban]

@HEB