Multi-agent: orchestrate & campaign
Add a fresh reviewer, split a goal into merged child runs, or fan one goal out to many gated orchestrators and compose the results.
Once a single gated run works, you can fan it out. The gate never changes: it just runs in more places. There are three escalating patterns.
review: a fresh set of eyes
orchestrate review runs a coder, then launches a
reviewer as an extend of the coder's run so
lineage is preserved. Point each role at a different CLI (say a Claude coder
and a Codex reviewer) and the reviewer verifies independently before
anything is accepted.
full-plan: split, run in parallel, merge
orchestrate full-plan asks a planner to split a goal into two to
six tasks, runs them as independent child runs (each gated on
its own), then merges them. The merge is dependency-aware: if
two tasks change the same file, the task that builds on the other wins, and
genuine conflicts are sent to a repair provider rather than dumped on you.
campaign: orchestrations of orchestrations
campaign lifts that one level higher: it splits a goal into
several full full-plan orchestrations and meta-merges
their results. This is the harness of harnesses at the top, a coordinator of
coordinators, with a hard depth cap of two and a tree-wide budget so it can't
run away.