Tom Hundley
Every week brings another headline about a foundation model that scored a fraction of a point higher on some benchmark, or a lab raising a round large enough to buy a small country's power grid. It is easy to mistake that noise for progress. Meanwhile, a British programmer who co-created Django two decades ago keeps shipping small, sharp tools that make AI genuinely useful against real data — and almost nobody in the executive suite is watching him closely enough.
They should be. On May 21, 2026, Simon Willison released the first alpha of Datasette Agent, which he describes as an extensible AI assistant for Datasette. It is not a billion-dollar launch. It is a 0.1a3 release. And it is one of the more instructive things to happen in applied AI this spring.
If you only know one thing about Willison, know this: he ships, and he writes down everything he learns while doing it.
His track record is not hype-adjacent. In 2003 he co-created Django, the Python web framework, alongside Adrian Holovaty at the Lawrence Journal-World newspaper. Django still runs a meaningful slice of the web. He later co-founded Lanyrd, went through Y Combinator, and joined Eventbrite through its acquisition, where he served as an engineering director. He has been blogging about web development and programming since 2002, sits on the board of the Python Software Foundation, and — a detail worth flagging for anyone thinking about AI risk — he is the person who coined the term prompt injection.
His current center of gravity is Datasette, the open-source tool he started as a side project to solve a data-publishing problem he had wanted to crack during his time at The Guardian. Datasette turns structured data in SQLite databases into explorable, queryable, publishable web applications. Around it he has built a constellation of smaller tools, most notably sqlite-utils for wrangling SQLite from Python, and since 2023, LLM — a command-line tool and Python library for talking to large language models from your terminal.
That LLM project is the connective tissue you need to understand the new release.
Datasette Agent merges Willison's LLM library into Datasette. The practical result is a conversational interface for your data: you ask a question in plain English, and the agent writes and runs the SQLite queries needed to answer it.
The design choices are the interesting part:
agent.datasette.io uses Google's Gemini 3.1 Flash-Lite, chosen because it is cheap, fast, and has no trouble writing SQLite queries. That is a quiet but important signal: useful data-agent work does not require the most expensive model in the catalog.datasette-agent-charts (visualizations via Observable Plot), datasette-agent-openai-imagegen (image generation), and datasette-agent-sprites (sandboxed code execution).The release did not stop there. In the days that followed, Willison shipped datasette 1.0a32 (May 31) and datasette-agent-micropython 0.1a0 (June 2), the latter adding an execute_micropython tool that runs model-generated Python inside a sandboxed MicroPython interpreter compiled to WebAssembly. That is a thoughtful answer to a real agentic-AI problem: how do you let a model run code without handing it your machine?
None of these carry a version number that would make a press release. That is exactly the point.
Willison's method is the opposite of the platform-team-first narrative dominating enterprise AI in 2026. He builds in public, ships small increments constantly, and writes up what he learns as he goes — to the point that his blog has become one of the most reliable practitioner records of the LLM era.
His view on writing is itself a useful executive lesson: he argues that consistent, frequent writing about what you are building compounds into outsized influence, precisely because so few people do it anymore. The output proves the thesis. When LLM 0.26 added tool support — letting a model in your terminal call any capability you can express as a Python function — Willison's framing was characteristically plain: tool use, he has argued, is the single most effective way to extend the abilities of language models. Datasette Agent is that conviction applied to the one thing every business actually has too much of: data sitting in databases nobody can easily query.
He is also candid about where he is taking it. He plans to use Datasette Agent to build a personal AI assistant — which, in a nice bit of foreshadowing for where this whole field is heading, he intends to call "Claw," wired into data imported from across his own digital life via his older Dogsheep tools.
Here is what I think the industry keeps missing.
The foundation-model race is real and it matters. But for the average organization, the bottleneck has not been model quality for a while now. The bottleneck is the unglamorous layer between a capable model and your actual data: how does a model safely query your database, run a chart, execute a snippet of code in a sandbox, and do it against a model you can afford to run a thousand times a day? That layer is plumbing. And Willison has spent three years building plumbing while everyone else argued about benchmarks.
There is a build-versus-buy lesson buried in here too. Datasette Agent is a counterexample to the idea that you need a large platform team and a seven-figure budget to ship agentic data tooling. One prolific builder, working in the open on top of SQLite and a model-agnostic abstraction, produced something a great many "AI platform" vendors are still pitching as a roadmap. The leverage a strong individual builder has in 2026 is genuinely disproportionate to their org-chart position — and that should reframe how executives think about both vendor selection and internal hiring.
I would not bet a production system on a 0.1a4 alpha tomorrow; Willison would not either, and he is refreshingly honest about that. But the direction is the bet worth watching. Model-agnostic, plugin-extensible, cheap-model-friendly, sandbox-aware, and documented in the open. That is what durable AI infrastructure looks like. It rarely trends. It just quietly becomes the thing everything else is built on — which is precisely the story of Django, and may well be the story of Datasette Agent.
What is Datasette Agent?
It is an extensible, open-source AI assistant for Datasette, first released as a 0.1 alpha on May 21, 2026. It lets you ask natural-language questions about data in SQLite databases; the agent writes and runs the SQL to answer them. A live demo is available at agent.datasette.io.
Which AI models does it work with?
It is model-agnostic. Datasette Agent supports hundreds of tool-calling models, including frontier APIs from OpenAI, Anthropic, and Google Gemini, as well as open-weight models you can run on your own hardware. The public demo uses Google's Gemini 3.1 Flash-Lite, chosen for being cheap and fast while still handling SQLite queries reliably.
Who is Simon Willison and why does his opinion carry weight?
Willison co-created the Django web framework in 2003, co-founded Lanyrd, was an engineering director at Eventbrite, and serves on the Python Software Foundation board. He created Datasette, sqlite-utils, and the LLM command-line tool, and he coined the term "prompt injection." He has blogged about software since 2002 and is widely read as a practitioner voice on LLMs.
How is Datasette Agent extended?
Through plugins, the same mechanism Datasette itself uses. Launch plugins included datasette-agent-charts for visualizations (via Observable Plot), datasette-agent-openai-imagegen for image generation, and datasette-agent-sprites for sandboxed code execution. A later datasette-agent-micropython 0.1a0 plugin (June 2, 2026) added an execute_micropython tool that runs Python in a WebAssembly-sandboxed MicroPython interpreter.
Is it production-ready?
Not yet — and that is by design. These are early alpha releases (0.1a1 through 0.1a4, alongside datasette 1.0a32). The significance is in the architecture and direction, not in deploying an alpha to production. Treat it as a strong signal of where agentic data tooling is heading rather than a finished product.
What should executives take away from this?
Two things. First, the value frontier in applied AI is increasingly in the "plumbing" — safely connecting capable models to your real data — not just in raw model benchmarks. Second, Willison's work is a concrete reminder that a single focused builder, working in the open on a model-agnostic foundation, can ship agentic tooling that rivals expensive platform roadmaps. Both points should inform build-versus-buy decisions in 2026.
Discover more content: