writing
field reports from the harness.
Architecture notes, postmortems, comparisons we actually checked, and the occasional unfiltered opinion about agent orchestration.
-
Running 50 Claude Code agents in parallel: what breaks first
When you scale from one coding agent to fifty, the bottleneck isn't the model. It's the orchestration layer. Here is what failed in our own experiments, in the order it failed, and how brat's design responds to each one.
operationsscale -
Designing crash-safe state for ephemeral agent processes
AI coding agents are processes that crash a lot — they hit rate limits, time out, get killed by their parent, or just decide to exit. If you treat their state as durable, you lose it. Here is how brat's append-only WAL design handles it.
architecturestate -
Why a thin CLI beats SaaS for agent orchestration
Multi-agent harnesses are heading toward hosted control planes by default. We picked the other path: a Rust CLI you install and forget. Here is the design reasoning, including the parts that are genuinely uncomfortable.
designopinion -
Mayor, Witness, Refinery, Deacon — the six-role vocabulary
brat names its orchestration roles after a small-town transport metaphor. It's a bit weird on first contact and very useful by the third week. Here is what each role does, why it exists, and what it explicitly does not do.
architectureroles