← ClaudeAtlas

nebi-orchestrationlisted

Read this before splitting work across sub-agents, not after. How the lead partitions and dispatches sub-agents so parallel work doesn't collide and sessions don't conflict on merge. Use when asked to split, fan out, parallelize, or divide work across agents; to dispatch, spawn, or assign builders, scouts, or checkers; when running Ultracode or another multi-agent workflow; when choosing how to run the fleet; or when work across sessions keeps conflicting. This applies even when the request is phrased as a direct command to start the work — load it first, then dispatch. Not for parallelism inside one file's code, such as concurrent calls, and not for CI workflow config.
romanized/nebi-skill · ★ 1 · AI & Automation · score 74
Install: claude install-skill romanized/nebi-skill
# Orchestration: route work without collisions The active harness's core override always wins, including any model, mode, or method Nebi named. Carry an explicit model choice through the dispatch tool when the harness exposes that control. ## Choose from dependencies and collision risk Map dependency direction, edit overlap, shared resources, and uncertainty before considering task size or agent availability. When the fan-out, loop, or ordering is knowable in advance, script it rather than letting a model re-decide each step. Use the least coordination that preserves clear ownership and proof: - **Solo:** one lead owns work whose files, decisions, or dependencies are tightly coupled. - **Scout:** one or more read-only workers gather heavy or independent evidence; the lead resolves contradictions and synthesizes. Unknown-size discovery repeats rounds until a round returns nothing new or a stated cap is reached. - **Serial workers:** hand off in dependency order when one output defines the next, or writers would share a file, resource, or decision. - **Parallel isolated writers:** use only with disjoint write scopes, satisfied dependencies, stable interfaces, and independent acceptance checks. - **Competing hypotheses:** assign read-only, falsifiable investigations; the lead or named arbiter chooses, then one owner implements. - **Generate and judge:** when the job is choosing among open-ended options (names, designs, rankings), generate candidates, then judge them against