Writing
Notes on harness engineering: the scaffolding around a language model — agent loops, tool schemas, context management, sandboxing, evals, telemetry — and the failure modes that only show up once something is actually running.
Most writing about agents stops at the happy path. I'm more interested in the layers underneath it, approached the way you'd approach any other system: bottom-up, one layer at a time, with attention to what breaks. Code for the series lives in a single repository that grows with each post rather than a graveyard of one-off demos.
First post going up shortly.