- Author
- Simon Willison (Simon Willison's Weblog)
- Published
- 14 March 2026
- Kind
- Personal blog post, highlights from a recorded talk
- Original
- simonwillison.net/2026/Mar/14/pragmatic-summit/
- Subject
- Harnesses and agent loops
- Themes
- Skill under automation, Handing over control
What it is
Simon Willison, a programmer who runs a weblog, wrote this post. It collects highlights from a fireside chat at the Pragmatic Summit in San Francisco. Eric Lui from Statsig hosted the chat. The post adds links to the author's earlier writing.
What it says
- Willison describes adoption stages, ending with not reading the code, which the talk calls irresponsible.
- Willison starts each agent session with the test command and a request for red-green TDD.
- Passing tests do not prove a server boots, so Willison has agents use curl.
- Willison built Showboat, which records an agent's manual tests as a markdown document.
- Willison describes conformance-driven development: tests that many frameworks pass guide a new implementation.
- Willison says code quality matters mainly for code maintained over the long term.
- Willison names the lethal trifecta and calls sandboxing most important, yet often runs Claude unsandboxed.
What evidence it gives
The post offers first-person experience and short examples. Examples include a file-upload test suite and a demo tool called Showboat. It gives no measurements or controlled tests.
Interests and limits
These are highlights chosen by the speaker, not the full conversation. The video was not viewed. All claims rest on personal experience. The author builds some tools mentioned, such as Showboat, and offers a paid newsletter.