David Gasquez

@davidgasquez

Data @ Protocol Labs. Open Data, Open Source, Open Protocols. Walks taker. Progressive Metal enjoyer. davidgasquez.com

1,348 Following5,372 Followers1,132 Posts

Bluesky

Is this your profile?Claim it and davidgasquez.dirty.social is yours.Claim profile

David Gasquez@davidgasquez· 7d

So cool! I tried with "a laptop" and got this awesome picture.

Screenshot of a British Library image viewer showing an 1873 illustration of a tilted book or stone slab surrounded by an arch of flowers. A dark metadata panel lists the French title, publication date, collection, scan link, and file path.

David Gasquez@davidgasquez· 7d

Wrote about why context engineering is a data problem. Organizations need to own the process that turns raw sources into useful artifacts for their agents. Much of this is work data teams already know how to do! davidgasquez.com/context-engi...

Companies want your data and context because [your context is their moat](https://x.com/samzliu/status/2080210797465379147).
Products that own your context can force you to use their agent... and, I've never seen a good hosted agent!

Organizations need to own their context, and the way to do so effectively already exists and is well understood.

## Context as Data Infrastructure

Have you noticed the [shape of the diagrams](https://storage.googleapis.com/gweb-cloudblog-publish/images/WorkspaceIntelligence.max-2200x2200.png) that every company is adding to their "intelligence" products? I see that and it seems like I'm looking at a Fivetran or dbt product page in 2018. That is because they're the same thing!

Most organizations will need to build and maintain a model-agnostic knowledge base ([aka ontology, company brain, ...](https://x.com/DBredvick/status/2078150905078206789)) for the same reasons they maintain a data warehouse. A curated and normalized layer helps both humans and agents make sense of all the structure that outlives any particular model, harness, tool, or product. Context engineering is [that same work](https://x.com/JoshARosen/status/2084693306722705629): extracting, filtering, curating, modeling, and publishing artifacts to help the organization make better decisions.

For now, it seems there isn't a great set of tools (or [modern context stack](https://x.com/davidgasquez/status/2082895547677954466)) built for this purpose. Something like Fivetran and dbt optimized for the new sources (Slack, Drive) and transformations (transcription, summarization, text extraction from a slide deck, ...).

David Gasquez@davidgasquez· 14d

One of the best things I did recently was to archive all @gordon.bsky.social's Substack posts into my personal digital library. I index, embed and query it with github.com/tobi/qmd. When my agents now $consult-library they stumble with many great Gordonisms, making them better for my taste!

Screenshot of a YAML configuration in a dark-themed code editor. It
defines a squishy-computer source with a local archive path, Markdown
glob pattern, commented Substack export command, and context describing
Gordon Brander’s essays on tools for thought, cybernetics, complex
systems, decentralized protocols, network society, and AI agents.

David Gasquez@davidgasquez· 15d

Also fun to see clusters like Bone Inmune Biology, Creative Culture Projects, and others that are not only Bluesky adjacent!

Zoomed-in semantic map showing interconnected clusters of colored links, including Alternative Internet Research, Bluesky User Profiles and Apps, AT Protocol resources, Tangled Code Repositories, Experimental Web Projects, Software Code Repositories, and Open Source Infrastructure.

David Gasquez@davidgasquez· 15d

Wanted to take a peek at @semble.so and didn't knew where to start. I built this as a static explorer of all the links in there. sites.wisp.place/davidgasquez... Love AT Protocol for letting me build the UI I want! 😍

Interactive “Semantic Field Map” for the Semble Link Atlas, showing
31,367 links grouped into 28 color-coded topic clusters. Thousands of
colored dots form a central map labeled with themes such as Experimental
Web Projects, AI Industry Criticism, Language Model Engineering, Books
and Literature, and Open Source Infrastructure; a searchable topic
legend with link counts appears on the right.

David Gasquez@davidgasquez· Jul 15

Write more ad-hoc ephemeral workbenches! They help you learn and a higher bandwith way to communicate with your agent. davidgasquez.com/ephemeral-ag...

When I work with agents, my first instinct is usually to “just prompt” through the task. That is fine for many coding-adjacent tasks, but not everything I do can be easily codified, and I bet the same thing happens to you!

I’ve explored how to make agents do subjective work better, and I think there is another underrated approach to using agents for subjective tasks, or tasks where the goals are fuzzy and you’d like to keep the final call: build local and ephemeral workbenches.

The idea is to have agents produce small pages/apps for the task at hand. For me, the clearest recent example is a grant review workbench I built. Instead of asking a model “which application is better?”, I got it to build me a workbench to help me better understand the task.

In the workbench, I could see and sort every application, add random tags or notes as I read through them, visualize them on a 2D map to spot duplicates or core applications, etc.

The useful part was the handoff. I could click a button to copy the state and ask my agent to do things like complete the labeling from my manual labels, recluster, add a new field to applications, suggest due diligence questions, etc.

This gave both of us a high-bandwidth way to share state and made the feedback loop faster than chat. Chat alone is a bad interface for this kind of work because too much state stays implicit, or even hidden from you. With the workbench, I could inspect both the applications and the changes the agent wanted to make.

David Gasquez@davidgasquez· Jul 13

Complexity is the mind killer. github.com/PhilipK/arti...

Complexity is the mind killer
TL;DR: When faced with a choice, always pick the simplest thing that solves your immediate problem. Then make sure you can change your mind later. By the simplest thing I mean the solution that is easiest to reason about.

When we always pick the simplest solution we avoid the analysis paralysis of finding out which solution might be the best in all the future scenarios we can imagine. We simply pick the simplest thing, and move on, knowing that it is easy to change our mind later. After we implement the simple solution, we have a much better understanding of our problem and we can ask "is my immediate problem solved?" If not; repeat.

Complexity is the mind killer. Our mind gets overloaded when we try to change a complex system, and this causes us to make mistakes. Systems must be changeable, it is what keeps them from becoming legacy and it keeps us agile and adaptable.

A lot of complexity gets added in the name of changeability. We create microservice systems so we can change our programming language and service implementation, but we add the complexity of network, service discovery, message queues, container orchestration and so on. It also adds a resource overhead that forces us to scale horizontally earlier, which is complex. Could we simply have hidden the implementation behind an interface or a function? Will we actually change our programming language any time soon? Does it actually solve our immediate problem?

David Gasquez@davidgasquez· Jul 1

Came across this old XKCD. Not only works for sports commentaty these days! 😅 xkcd.com/904

Two stick figures talk. Caption: “A weighted random number generator just
produced a new batch of numbers.” One says, “Let’s use them to build
narratives!” Bottom caption: “All sports commentary.”

David Gasquez@davidgasquez· Jun 19

Got 3.3M links (follows of everyone I follow on Bluesky) into a cute map. These are the bubbles of my feed. Data people, AI people, ATProto builders, OSS, ... sites.wisp.place/davidgasquez...

David Gasquez@davidgasquez· Jun 12

Not sure how well am doing it (feedback appreciated) but from now on my posts are also published to the Atmosphere via @standard.site! 🌱 Tiny sync script in case you want to steal it. 👇 github.com/davidgasquez....

all my posts as they appear on davidgasquez.com

David Gasquez@davidgasquez· Jun 11

Cool to see @hypercerts.org on the What's Hot list of community lexicons! 🤩

David Gasquez@davidgasquez· Jun 3

I keep using this qmd setup for research and it is really great! The workflow is simple: put a bunch of resources in a YAML file and qmd will pick them up and index/embed them. Answers are more grounded than just web search and I can control more what the models see!

Screenshot of a dark-themed code editor with a YAML index file open
beside a Markdown spec titled “Atmospheric Data Portals.”

David Gasquez@davidgasquez· May 29

Content defined chunking and contend addressed storage rocks indeed!

Screenshot of a tweet by Julien Chaumond saying Hugging Face is becoming bullish on data infrastructure. The tweet says he cloned 68 TB to a Hugging Face training bucket in under two minutes using Xet deduplication and infrastructure optimizations. Below the text is a video still of a Hugging Face “Copy to bucket” dialog showing a copy in progress from jasperai/monet to julien-c/my-training-bucket/monet-copy, with 178+ files, total size 68.2 TB, and an estimated time of less than one minute.

David Gasquez@davidgasquez· May 23

Wrote a small post advocating for more mindful sharing of raw AI outputs. I'm feeling it more and more (specially at work) and would love to reiterate it once more in writting.✌️ davidgasquez.com/keep-your-sl...

You’ve probably heard this already but, I keep coming across this pattern and I wanted to add another post to the cause.

Generating a wall of text is now free while reading, verifying, and distilling still cost time and effort on the recipient. That asymmetry is what makes you sharing that document consisting of raw unrequested AI output rude. As this vibecoded website says, stop sloppypasta!

The overall principle I keep in mind is what you send should take you more effort to produce than it takes me to read. This applies to chats, but also to code, bug reports, PRs, emails, docs, and now, it seems slides too.

This is also especially bad when the point of the writing is to demonstrate your thinking. Rewriting that with an LLM changes meaning subtly, blurs authorship, and erodes voice. And people can tell.

There is also “good uses of slop” though! I think it is ok to:

Send a draft you’ve read, edited, and would defend as yours.
Share the output as a side artifact. I like to see other folks’ raw prompts and sessions (I’ve learned a lot from reading Simon Willison prompts) so I can learn from it or fork it and tweak it.
So, two small asks:

Keep slop to yourself and share drafts once you’ve read it, verified it, and distilled them down to what actually matters.
If you do share it, disclose it. Ideally send the prompt and a link to the chat rather than the output. Sessions are more interesting than transcripts, and the recipient can tweak and run their own models on top of it (I don’t usually trust people’s LLM context management skills). If your prompt is too embarrassing to share, that’s a signal worth listening to.
These new manners are still emerging and I’m probably also offending folks with this sometimes (slopiness is an spectrum). But, let’s at least try not to outsource our thinking onto each other. Otherwise, we all end up drowning in slop or delegating to agents to get more slop!

David Gasquez@davidgasquez· May 18

Wrote about why the AT Protocol is the ultimate API for you and your agents! davidgasquez.com/atproto-agents We need more people building on the atmosphere!

My atmosphere data lives at at://davidgasquez.com. You can explore it all without API keys, auth, or arbitrary HTML to parse. Your agent can browse it, query it, and link to it. Getting someone latest posts is one click/curl away.

That’s the pitch. The rest of this post is why it works, and why it might be the substrate your agents have been waiting for.

Hostile Platforms
You’ve probably experienced more than once recently, platforms being hostile to you (or your agent) getting the data out: throttled or limited APIs, bans for scraping, no identity persistence, no real-time access, walled gardens. Every integration is a deal that can be revoked when the CEO wakes up in a bad mood.

The AT Protocol is an amazing technology that I think is very underrated for agents! Turns out, the properties Bluesky (the biggest AT Protocol application as of today) needed for humans (portable identity, open data, federated infrastructure, structured schemas) happen to be exactly what you’d want for agents too.

As a quick and dirty intro (you should read the official ones, or this great walkthrough by mackuba), in the atproto world, every user is a personal, signed JSON repository. Every record (a post, a like, a follow, a photo, a blog post, anything really) is addressable by an at:// URI, has a public schema (a Lexicon), and is broadcast in real time on a global event stream (the firehose). Dan Abramov explains all of this in more detail and clarity in “A Social Filesystem”.

David Gasquez@davidgasquez· Apr 23

Two things I've done am very happy with: - Publishing a SKILL.md file with references to datasets docs, code, and other useful context (how to plot) - Index anything a data engineer could need. Core business logic docs, external repositories, APIs, ...

David Gasquez@davidgasquez· Apr 23

Two months later, happy to report the barefoot data platforms approach has been working quite well! Shared some learnings on a small post. davidgasquez.com/growing-my-o...

It has been a couple of months since I wrote about Barefoot Data Platforms. Since then, I’ve been applying those ideas into a real data platform I’ve been rebuilding from scratch. Learned a lot about what makes these platforms work well, especially with agents.

Building It
As a quick overview, the entire orchestration engine (fdp) is ~700 lines of Python. Assets are plain .sql or .py files with ugly but useful metadata in comments. No decorators, no registration, no config files. You can drop a file in assets/, and the next run picks it up, resolves dependencies, and materializes it! Delete the file, and it’s gone too.

This minimalist, low-abstraction and filesystem based conventions data platform is great fit for agents:

David Gasquez@davidgasquez· Mar 24

In earlier eras of computing unreliable hardware forced system design discipline (checksums, retries, modularity). There is a lot of that we are rediscovering in the "Agentic Engineering" era! Wrote a small post about this idea. davidgasquez.com/reliable-unr...

Agentic engineering is mostly about building reliable systems around unreliable components (like your friendly coding agent).

A good analogy I like is how early computers were powerful, but not trustworthy enough to be used “raw”. Hardware failed, bits flipped, and storage was noisy. Engineers had to build systems around the machine to make things more predictable. And, they came up with a bunch of interesting ideas!

Error correction codes
Redundancy
Checksums
Validation layers
Retry logic
With these, even though reliability was not a property of the machine/computer, it became a property of the system.

This is something close to where we are with “Agentic Engineering”.

As you’ve probably experienced, a coding agent can be fluent, useful, and very wrong at the same time! The key is to, like the early programmers did, treat agents as noisy components instead of trying to get the perfect/bigger one.

David Gasquez@davidgasquez· Mar 19

Made a silly CLI to run queries over markdown files and their frontmatter inspired by Obsidian Bases. github.com/davidgasquez...

Terminal running through the tool examples.

David Gasquez@davidgasquez· Mar 18

New wallpaper perhaps?

A person in flowing white, sheer fabric stands on a stone balcony inside a large circular architectural space, their gauzy garments billowing dramatically as if caught by wind. Behind them, a tall narrow doorway opens into darkness, framed by concentric textured patterns that emphasize the scale and symmetry of the structure.

David Gasquez@davidgasquez· Mar 12

Built a tiny pi.dev extension so I can launch task-specific agents. Each profile is a small YAML file: model, thinking level, system prompt, and allowed skills. Simple, contained, and surprisingly useful! davidgasquez.com/specializing...

David Gasquez@davidgasquez· Feb 20

Cleaned and updated my notes on Agentic Engineering. davidgasquez.com/handbook/age...

An agent runs tools in a loop to achieve a goal. Agents are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks. Agents help decouple programming, the craft of physically typing code, from engineering, the architecture of your system, the goals, the “why” of what you’re building.

Using LLMs for coding is difficult and unintuitive, requiring significant effort to master.
Don’t delegate thinking, delegate work.
Before coding, make the plan with the model.
You can use the same or a different model to critique the plan and iterate. If you are unsure, ask to “give a few options before making changes”.
Redoing work is extremely cheap. Prioritize exploration over execution (at first). Iterate towards precision during the brainstorming phase. Start fresh once you know what and how to build it.
Failed attempts are cheap. If the plan fails and the result is bad, just delete everything and try again.
Divide the problem into smaller problems (functions, classes, …) and solve them one by one. Keep sessions short.
Use Progressive Disclosure to ensure that the agent only sees tasks or project-specific instructions when it needs them.

David Gasquez@davidgasquez· Feb 11

Half baked idea and proof of concept but at least it is out there. I wrote about Tributary, a credibly neutral mechanism to incentivize and elicit useful datasets. davidgasquez.com/tributary-da... I worked on it 6 months ago at a research retreat (IERR 2025) but never wrote about it.

David Gasquez@davidgasquez· Feb 10

Also made a cute embedding of all their products! datania.github.io/mercadona-ca...

David Gasquez@davidgasquez· Feb 10

Discovered Mercadona's API and wrote a quick scrapper that will run weekly and dumps everything into JSONs in HuggingFace. Useful to track prices over time, check new arrivals, ... Data. huggingface.co/datasets/dat... Code. github.com/datania/merc...

David Gasquez@davidgasquez· Feb 9

Spotted this interesting design/implementation of a self-hosted context store for agents. - Turn DAG - Blob Deduplication / Content Addresses Storage - BLAKE3 - Dynamic types, visual debugging, ... github.com/strongdm/cxdb

CXDB - AI Context Store
CXDB is an AI Context Store for agents and LLMs, providing fast, branch-friendly storage for conversation histories and tool outputs with content-addressed deduplication.

Built on a Turn DAG + Blob CAS architecture, CXDB gives you:

Branch-from-any-turn: Fork conversations at any point without copying history
Fast append: Optimized for the 99% case - appending new turns
Content deduplication: Identical payloads stored once via BLAKE3 hashing
Type-safe projections: Msgpack storage with typed JSON views for UIs
Built-in UI: React frontend with turn visualization and custom renderers

David Gasquez@davidgasquez· Feb 4

Pasted an screenshot of the conversation and got this done! I don't know if you get used to this feeling. 😅 davidgasquez.github.io/barefoot-dat...

-- asset.name = filtered_numbers
-- asset.schema = raw
-- asset.description = Filtered values from transformed numbers
-- asset.depends = raw.transformed_numbers
select
value,
square,
label,
double,
parity
from raw.transformed_numbers
where value >= 3

David Gasquez@davidgasquez· Feb 3

Wrote about "Barefoot Data Platforms", my take on minimal and agent friendly data platforms! davidgasquez.com/barefoot-dat...

So, how does a minimal and opinionated data platform look like? Minimalism for me here means low abstraction and no frameworks. You write your scripts. These scripts are assets, and these assets depend on each other 1. That’s it. The opinion side of things is the interesting part. Here are the key ideas of the Barefoot Data Platform I built:

Classic functional, declarative, independent, composable, and idempotent transformations
Co-located assets, metadata, and documentation so assets have everything needed to be understood and reproduced in the same file
One storage layer. DuckDB or SQLite, Parquet files, or whatever you want as long as you define a materialize function
Many automated checks (ruff with all the rules enabled, ty, data tests, prek hooks, custom bdp check, …) for the agents

David Gasquez@davidgasquez· Jan 24

TIL you can install Python tools without uv or Python with a single curl command! uvx.sh Useful for Docker, CI pipelines, or anywhere you don't want to mess with the Python installation or venvs!

uvx.sh
Install Python tools with a single command. Powered by uv.

For example, to install ruff:

David Gasquez@davidgasquez· Dec 19

Updated the design of my personal website and I love it! davidgasquez.com/handbook