Agent System Format
A private-work-informed project direction for describing agent swarms with roles, traces, artifacts, context, and evaluation.
This is the current main project direction behind the homepage: a format for making multi-agent workflows concrete enough to run, inspect, and improve.
The public description is intentionally high-level. The useful parts to expose are the ideas: assignment manifests, role boundaries, execution traces, artifact-first persistence, context pieces, and judges that decide whether a run actually succeeded.
The private implementation details stay out of this site, but the blog can still publish reusable design notes around the format.