- Pet agents
- Interactive coding-agent sessions you name, nurture, and debug individually — high-touch, high-context, one at a time. The counterpart to cattle agents. Read the note →
- Cattle agents
- Headless, replaceable agent workers managed as a herd: when one fails you don't debug the worker, you investigate the task and dispatch another. The unit of analysis is the fleet, not the session. Read the note →
- Bead
- A single unit of work in a file-backed queue (the beads/br tooling): title, description, acceptance criteria, and dependency edges, claimable atomically by exactly one worker. Read the note →
- Genesis bead
- The root tracking bead for a whole project: it references the plan document, carries the phase checklist, and closes only when every phase bead beneath it is done. Read the note →
- Strand escalation
- What a worker does when its queue is empty: search other workspaces for claimable work, do cleanup, propose alternatives — an empty queue is a signal, not idle time. Read the note →
- Stitch
- A proposed new piece of work surfaced by an observer (like HOOP's human-interface agent) that only becomes a bead after operator approval.
- Outcome handler
- The rule that if an outcome can happen, it has an explicit handler — no wildcards, no swallowed errors. The design constraint that makes unattended fleets runnable. Read the note →
- The plan is the prompt
- Treating the plan document as the token-dense artifact a headless fleet executes from: explicit constraints, per-task acceptance criteria, and verification commands instead of conversational context. Read the note →
- The agentic coding ladder
- A reframing of the levels of agentic coding by what you give up at each rung — the price of each level of automation, not just its capability. Read the note →
- Grading its own homework
- Letting the agent that produced work also judge that work. The fix: independent verification — a live artifact, a second agent, or a different tool. Read the note →
- Anchor fact
- A ground-truth fact (usually a number) deliberately withheld from the model, used afterward to audit whether its output can be trusted. Read the note →
- Ending over mending
- When agent output misses, reverting and re-dispatching is usually cheaper than hand-fixing — the economics of regeneration versus repair. Read the note →