Behind the scenes of building Savant — the challenges, the bugs, and what we learned.
July 24, 2026
A tool call hung for three minutes. The JSON-RPC response never arrived on stdout, but the thing that killed it was a log line on stderr. Anatomy of a pipe-backpressure deadlock in an MCP transport.
Read more →July 17, 2026
A user asked for scripts that can run pandoc. I spent three months writing threat models about it, then went to check the five permissions I'd already shipped. Three had holes. The other two didn't need any.
Read more →July 3, 2026
The Claude API can fail inside a stream that returned HTTP 200. The parser had no handler for the error event, so users got empty replies saved as complete. Anatomy of a failure where every layer reported success.
Read more →July 2, 2026
A realistic 7-turn agentic conversation billed 219,785 full-rate input tokens to produce 34,000 tokens of output. The cache looked enabled, the accounting said otherwise, and the fix that mattered was a sort and a loop.
Read more →