Cost-tiered multi-agent delegation for Claude Code: a Fable or Opus master orchestrator fans out a swarm of cheaper, specialized agents only when it pays.
Agent
Writes production code where a mistake is expensive. Dispatch for real feature work, non-trivial bug fixes, refactors that must preserve behavior, and anything touching correctness-critical paths. Deliberately runs on the top model tier — do not route work here that the mechanic agent could do from a precise spec, and…
Agent
Build-workflow leaf. Fills the bodies of pre-written signature files for one manifest unit or a small batch of them. Dispatch from the claude-swarm:build workflow, not ad hoc — it expects exact owned paths, fixed signatures, and a type-check command. Writes only the files its unit owns, runs the type check before…
Agent
Applies a precisely specified change across a codebase. Dispatch for mechanical work where the decision is already made and only the execution remains — apply this pattern to these N files, add these headers, rename this symbol everywhere, update these call sites to the new signature. Cheap because the compiler and…
Agent
Cheap read-only locator. Dispatch when you need to find WHERE something is — the file defining a symbol, every call site of a function, which package owns a concept, whether a pattern exists anywhere in the repo. Returns paths and line numbers, not explanations. Use several in parallel to sweep a repo from different…
Agent
Writes and maintains prose about code — README sections, docs, changelogs, release notes, code comments, config examples. Dispatch when the content is documentation rather than implementation. Cheapest agent in the roster; use it freely for writing tasks and never for deciding what the code should do.
Agent
Deep reader that returns a short answer. Dispatch when understanding something requires reading a lot of code — how a flow is wired end to end, what a subsystem actually does, why two components interact, what the real control flow is behind an abstraction. Reads thousands of lines and returns a few hundred words. Use…
Agent
Adversarial checker. Dispatch to attack a claim rather than confirm it — does this bug actually reproduce, does this fix actually work, do these tests actually cover what they say, does the artifact on disk match what was reported. Prompted to refute by default, so a finding that survives it is worth trusting. Run…