Scope Sizing for Agents

Scope sizing is the art of making each task unit the right size for an agent. Too small, and the coordination overhead exceeds the task itself — you spend more tokens on hand-offs than on work. Too large, and the agent loses focus, exceeds its context window, or produces lower-quality output because it is juggling too many concerns simultaneously. The sweet spot is a task that takes 5-15 minutes of agent time and produces a coherent, testable artifact.

Three signals indicate a task is too large for a single agent: it touches more than 8-10 files, it requires reading more than 2000 lines of existing code for context, or it has more than 3 distinct subtasks that could be individually tested. Three signals indicate a task is too small: the hand-off document is longer than the task output, the setup instructions exceed the implementation, or the task can be expressed in a single sentence without losing precision.

# Scope sizing heuristics

TOO SMALL (merge into larger task):
  ✗ "Add a single import to config.ts"
  ✗ "Rename variable from x to y in one file"
  ✗ "Add a comment to the auth middleware"
  → Hand-off overhead > task value

RIGHT-SIZED (ideal agent unit):
  ✓ "Implement rate limiting middleware with config + tests"
  ✓ "Refactor the auth module from callbacks to async/await"
  ✓ "Write integration tests for the payment flow"
  → 5-15 min agent time, testable output

TOO LARGE (decompose further):
  ✗ "Rewrite the entire API layer"
  ✗ "Add authentication, authorization, and audit logging"
  ✗ "Migrate from REST to GraphQL"
  → Multiple distinct deliverables, context overflow risk