Part 2 of 6 · 2 min read
Definitions and setup
Covers: Definitions; The decision depends on task shape.
Definitions
Sub-task, dependency, level, edge
- Sub-task: a unit of work that one agent can complete in a few steps.
- Dependency: sub-task B needs a result produced by sub-task A.
- Level: the length of the longest chain of dependencies; with all sub-tasks independent there is one level.
- Edge: one dependency, which means one fact that must be handed over.
Fan-out, orchestrator, worker
- Fan-out: dispatching independent sub-tasks to several agents at once.
- Orchestrator: the agent that plans, dispatches and combines.
- Worker: an agent that executes one sub-task in its own context (module 8 of Level 7 shows what isolation does to cost).
Shared-context tax, redundancy, critical path
- Shared-context tax: the background (task description, repository summary, policy) that every worker must load into its own context, so that it is paid once per worker rather than once.
- Redundancy: extra work done because a worker lacks the whole picture, such as repeated searches.
- Critical path: the longest chain of calls that cannot overlap; it sets the wall-clock time.
The decision depends on task shape
Three accounting facts carry the argument:
- Latency is levels, not sub-tasks. Parallel workers finish in the time of the slowest chain. A chain of eight costs as much time as one agent doing all eight, plus the relays.
- A single agent re-reads its growing context (module 8 of Level 7), so its cost rises with the square of the steps unless a cache makes the re-reading cheap. Workers keep contexts short, which is the saving; the shared-context tax and redundancy are the cost.
- Each edge is a handoff. The worker must be told the upstream fact, and a brief can lose it; the single agent has it in context.