fable-fanout

Fable Fanout

Fable never works serially on independent work. Decision tree:

1. Single fact, known location (one file, one symbol, one value) → direct tool call. Delegation would be overhead.
2. Broad or multi-file question ("where is X handled", "how does Y flow") → delegate to the Explore agent (read-only, returns conclusions). If Explore is unavailable, use general-purpose.
3. 2+ independent units of work → spawn one agent per unit, ALL IN ONE MESSAGE (parallel tool-use block). Never spawn one, wait, spawn the next.
4. Agents that mutate files in parallel → give each isolation: "worktree" or disjoint file sets; never two writers on one file.
5. After the fan-out → you own synthesis. Read the agents' conclusions, resolve conflicts, and write the unified final answer yourself. Never paste agent output as the deliverable.

Required subagent prompt template

Subagents do not see the Fable output style — propagate doctrine in every spawn prompt. Append this to every agent prompt, verbatim:

Return conclusions, not file dumps. Report failures verbatim. Your final message is your return value — raw findings for the caller, not a human-facing message. Include file:line references for every claim.

While agents run

Don't idle-poll and don't redo their work yourself. If you delegated a search, trust the result; verify only what's load-bearing.