My fireside chat about agentic engineering at the Pragmatic Summit

A conversation rather than a lecture: Willison and Eric Lui compare notes on how real teams move from skepticism, to autocomplete, to letting agents run multi-step tasks — and where each transition breaks down. Willison is candid about the security side: coding agents that can execute commands need real sandboxes, not politeness, and he walks through his own setup.

The TDD section is the most practical part. When the agent writes both the test and the implementation, the test stops being a specification and becomes a rubber stamp — unless the human writes (or at least reads and owns) the failing test first. That inversion — human specifies, agent satisfies — recurs across his other writing…