datasette-agent 0.3a0

The more consequential detail buried in this release is the approval-gate architecture itself. Willison is essentially shipping a mini-governance layer inside a single plugin, which raises the question of whether that pattern should live at the agent level, the database driver level, or somewhere in between.

This release sits in a different part of the stack than most of what we've covered recently. The OpenAI Codex iOS integration from June 15 is the closest parallel in spirit: both stories are about collapsing the distance between AI-assisted editing and actual execution against a live environment. But where Codex's iOS work focuses on read-and-preview workflows, datasette-agent is explicitly crossing into mutation territory, which is the harder trust problem. The related coverage doesn't map cleanly here because datasette-agent is an independent open-source project, not a platform play from a major lab. The relevant comparison class is other human-in-the-loop agent frameworks (LangGraph, Semantic Kernel) rather than anything we've directly covered.

Watch whether Willison ships a configurable approval policy (row-level, table-level, or operation-type-level) in the next two or three releases. If he does, that signals datasette-agent is targeting production deployments seriously rather than staying a developer toy.