The plugin extension model is the underrated piece of this release. Datasette Agent ships with three plugins on day one (charts, image-gen, sprites), which means the agent surface is being designed for third-party tool integration rather than as a closed product, and that is the inversion of how most chat-first AI agents launch today. Three years of LLM library work finally meeting the SQL grounding is what makes this a different shape of agent: queries are auditable, plugins are inspectable, and the data layer is the stable substrate underneath.
Model-agnostic design reads as an architectural choice, but given the pricing section of this same post, it's also a financial hedge. Flash approaching Pro pricing, GPT-5.5 doubling over 5.4, Opus 4.7 at 1.46x: when your agent layer treats inference as interchangeable, you can arbitrage between providers as they probe price tolerance. Value locks into the plugin and data access layer instead. That uv one-liner switching between Gemini and local gemma-4 is doing more strategic work than it looks.
Was the surprising part here that the cheap/fast model was already good enough for SQLite? I’d be curious where it first breaks: schema discovery, join planning, or quietly writing a very plausible query against the wrong table.
The pelican sighting demo proves Gemini 3.1 Flash-Lite writes accurate SQLite queries without needing expensive reasoning models.
I track this builder migration toward lightweight retrieval tools at theaifounder.substack.com.
What prompt construction techniques do you use to filter irrelevant schema data before the agent attempts a query?
The plugin extension model is the underrated piece of this release. Datasette Agent ships with three plugins on day one (charts, image-gen, sprites), which means the agent surface is being designed for third-party tool integration rather than as a closed product, and that is the inversion of how most chat-first AI agents launch today. Three years of LLM library work finally meeting the SQL grounding is what makes this a different shape of agent: queries are auditable, plugins are inspectable, and the data layer is the stable substrate underneath.
Model-agnostic design reads as an architectural choice, but given the pricing section of this same post, it's also a financial hedge. Flash approaching Pro pricing, GPT-5.5 doubling over 5.4, Opus 4.7 at 1.46x: when your agent layer treats inference as interchangeable, you can arbitrage between providers as they probe price tolerance. Value locks into the plugin and data access layer instead. That uv one-liner switching between Gemini and local gemma-4 is doing more strategic work than it looks.
Was the surprising part here that the cheap/fast model was already good enough for SQLite? I’d be curious where it first breaks: schema discovery, join planning, or quietly writing a very plausible query against the wrong table.