Boyd’s Agent turns Claude Code into a persistent strategic analyst. It maintains your orientation and your adversary’s across sessions, surfaces the mismatches between what you expected and what happened, and runs full OODA cycles on demand — orientation first, not speed first. No prior Boyd knowledge required: the doctrine ships with the agent.
The point is not to cycle faster. The point is to have a better, more coherent, more rapidly updated Orientation than the adversary.// FROM THE PROJECT’S FOUNDING ESSAY
The popular OODA loop is four boxes in a circle and a moral about speed. Boyd’s last diagram (The Essence of Winning and Losing, 1995) is something else: Orientation shapes which observations register at all, which options surface during decision, and which actions are even considered. There is even an implicit-guidance bypass from Orientation straight to Action.
This agent is built around that harder version. It does not chase tempo for its own sake. It builds and re-builds your orientation — and your adversary’s — and makes the structure of your situation legible enough to act on. Speed follows from that, when it follows at all.
This is not a chatbot with a strategy persona. The structure is the product.
Each system lives in its own directory with a documented schema: orientation.yaml for your side, adversary.yaml for theirs, an append-only observation log, and an open/closed list of mismatches. Across sessions and weeks, the picture compounds.
Reasoning is split across three subagents with non-overlapping write permissions. A Blue orienter holds your view; a Red orienter holds the adversary’s. They run in parallel so your model of the other side isn’t contaminated by your model of your own.
The agent does not act on the world. It produces a next action with rationale, branches, and sequels — and stops. The seam between proposal and execution is preserved as a feature, not a limitation.
Falsifiable mental models with explicit confidence and evidence. Each cycle runs Destruction & Creation on at least one of them — pulling a model apart and recombining it against the latest observations.
Reasoned from the adversary’s perspective using only what we can plausibly know. Unknowns are marked as unknowns. The Red orienter never writes to Blue’s files, and never hallucinates their internals.
Coordination between them comes from a shared Schwerpunkt — a focus of effort, not a target.
The adversary doesn’t have to be a company — the repo ships worked examples where it’s a hiring process and a procurement org. More in the FAQ.
Install from the marketplace, then bootstrap your first system. Three worked examples ship with the repo — a SaaS company facing a funded incumbent, a job search, a contract negotiation — so you can read a real one first.