paralelizacao-em-grafo

First close the contract (schema, types, error format), then split work into vertical slices where no file is read or written by two slices. Run each slice in its own git worktree with its own verification gates, and finish with a single-owner fan-in.

Avoid parallelizing single-task waves, slices sharing files, work with an open contract, tasks depending on conversation history, or when token cost matters more than deadline. In these cases sequential execution or one owner for the coupled set is cheaper.

A vertical slice gives one agent an entire feature (schema, API, UI) with exclusive files, so pieces join cleanly at fan-in. A horizontal slice splits by layer (models, routes, types), forcing agents to share files and turning the merge into a negotiation.

git worktree gives each agent a separate checkout of the same repository, preventing conflicts or silent overwrites when slices edit concurrently. Create one worktree per slice with its own branch, then remove the temporary worktrees after fan-in.

Each parallel window re-reads the contract and codebase from scratch, so N slices cost N times the prefill tokens, drawn from the same subscription quota. The clock time saved is only the slowest slice, while the bill multiplies.

The integrator reviews all slice outputs in one window, checks integration seams like cross-layer mismatches, and signs off a single result. The integrator must not implement code, and failed slices are returned to their worktrees rather than merged.