B8.1Bloom · RNot started

Two valid spawn triggers

Reading depth

What you'll learn

Spawn a subagent only for parallel work or to keep messy intermediate context out of the parent. Cost, window size, and tool count are NOT reasons.

  • Two and only two triggers: parallelism, isolation.
  • Parallelism: independent work, fan-out reduces wall-clock latency.
  • Isolation: intermediate context kept out of parent.

There are exactly *two* valid reasons to spawn a subagent: **parallelism** (work that can run independently and concurrently) and **isolation** (work whose intermediate context would otherwise pollute the parent). The exam tests this directly and frequently uses 'cost', 'window size', 'tool count threshold', and 'feels cleaner' as plausible-sounding distractors. None are valid triggers.

**Parallelism**: 'fan out 12 independent KB searches' or 'review 6 PRs at the same time'. Each unit of work doesn't depend on the others; running them sequentially in a single agent loop is N× slower with no benefit. Coordinator-subagent shape: parent dispatches, each subagent does one unit, parent integrates.

**Isolation**: 'run 30 turns of exploratory debugging without bloating the parent's planning thread'. The subagent's intermediate scratch — failed hypotheses, large tool outputs, dead-end diffs — stays inside its context. The parent only sees the final summary. The parent's context stays focused on the high-level plan.

The non-triggers worth memorising as distractors. Subagents are **not cheaper** (each is its own model invocation chain). Subagents do **not have a bigger context window** (they have a *fresh* one of the same size). There is **no tool-count threshold** that mandates subagents. 'Feels cleaner' is not a structural reason — splitting trivially-sequential work just burns tokens. The exam will offer all four; the right answer is parallelism or isolation.

Key points

  • Two and only two triggers: parallelism, isolation.
  • Parallelism: independent work, fan-out reduces wall-clock latency.
  • Isolation: intermediate context kept out of parent.
  • Non-triggers: cost, window size, tool count, 'cleaner' — common distractors.
  • Coordinator-subagent shape: parent dispatches + integrates; subagents do the heavy work.

Examples

Parallelism: 12 KB searches

Each search is independent. 12 sequential searches in one agent = 12× the time + parent context grows with each search trail. 12 subagents in parallel = wall-clock reduced to ~one search; parent context only carries the integrated summaries.

Isolation: exploratory debugging

30 turns of trying things, most of which fail. Inline: parent context fills with dead-ends and the plan thread is buried. Subagent: dead-ends stay isolated; parent only sees 'I narrowed it to module X, here's the patch'.

Pitfalls

  • Reaching for subagents for cost reasons. They cost more.
  • Reaching for subagents because 'I have a lot of tools'. There's no threshold.
  • Reaching for subagents to 'organise' sequential work. Sequential work is for the parent loop.
  • Confusing 'fresh window' (true) with 'bigger window' (false).

Source notes: 00-academy-basics/notes/08-subagents.md

Ask Claude

Build a prompt with this lesson + your question, copy it, and open the Claude Project in a new tab.