subagent-brief

Subagent Brief Discipline

Pre-flight checklist for spawning subagents in Claude Code. The single most expensive mistake in multi-agent workflows is treating subagents as free threads. They are not. Each subagent is an independent LLM call that re-tokenizes the entire prompt you hand it.

There is no prefix sharing across subagents in Anthropic's serving stack today. The multi-agent topology is a known optimization opportunity (see arXiv 2604.25899 Pythia: Predictability-Driven Agent-Native LLM Serving, 2026-04) but has not yet shipped in production. Until it does, the burden of compression falls on the orchestrating agent.

The math: 20 subagents × 50K shared context = ~1M tokens per fan-out. One careless Task call exhausts a Max-plan daily quota.

When to Use

• About to invoke the Task tool or any subagent-spawning tool
• About to fan out N parallel subagents on similar tasks
• Subagent prompt is shaping up longer than ~200 words
• About to embed a long file, full repo dump, or full PRD in a subagent prompt
• The user is on a metered plan (Max / Pro / API) and has not signaled "burn budget"

When NOT to Use

• Spawning a single subagent with a one-line task ("read X and summarize")
• The subagent genuinely needs the full document (translation, full audit, line-by-line review)
• Throwaway exploration where token cost truly does not matter
• The orchestrator already received a compressed brief from the user — pass it through, do not re-compress

The Discipline

For every subagent you are about to spawn, apply these five rules in order.

1. Strip your own context

The subagent gets its own system prompt, its own tool list, and its own task framing automatically. Do not paste your own conversation history, your TODO list, or your reasoning trail into the subagent prompt unless the subagent literally needs them to do its job.

2. Replace files with paths

BAD:  Here is the full content of foo.py: <800 lines>
GOOD: Read /abs/path/to/foo.py and report the public API.

The subagent has Read access. One Read call is cheaper than re-tokenizing 800 lines on every spawn.

3. Replace context with summaries

BAD:  <entire 50K context window dump> — now look for X
GOOD: Project context: <3-line summary>. Look for X in <specific path>.

If you genuinely synthesized something useful from your context, summarize it in 1-3 sentences and pass that summary. Never paste the raw window.

4. Compress to ≤200 words

The brief should fit in a tweet-length window. If it does not, you are either dumping context the subagent can fetch itself, or handing over a task too large for one subagent — split it.

5. Fan-out check

Before spawning N parallel subagents, ask: do they all share the same long context? If yes, you are about to multiply cost by N. Either:

• Spawn 1 subagent that handles all N items sequentially, or
• Pre-compute a shared brief once and hand each subagent only the delta (the per-item task), not the shared context.

The Brief Template

ROLE:    <one line — what this subagent is>
GOAL:    <one line — what it should produce>
INPUTS:
  - <path or 1-line summary, never raw content>
  - <path or 1-line summary>
CONSTRAINTS:
  - <output format, what NOT to do>
RETURN:  <one line — what to report back>

Total ≤ 200 words. If it does not fit, the task is too large for one subagent. Split it.

Rationalizations to Reject

Excuse	Reality
"Faster to just paste the whole file."	Faster for you. The subagent re-tokenizes 800 lines on every spawn.
"Parallelism will be 5× faster."	5× faster wall-clock, 5× more tokens — and Max plan caps are daily, not hourly. The hourly speedup costs you the rest of the day.
"The subagent needs context."	It needs information. A summary is information. The raw window is dump.
"I'll just pass the whole repo, it's small."	"Small" repos compound across 10 subagents into a Max-plan kill.
"I don't have time to summarize."	Summarizing forces you to clarify what the subagent should do — which is 80% of why subagents fail anyway.
"The user wanted parallel."	The user wanted fast. Token-budget parallel ≠ wall-clock parallel. Default to sequential unless tasks are genuinely independent.

After the Brief

Spawn the subagent. Then evaluate:

• Returned with hallucinated assumptions → brief was too vague. Add a constraint and re-spawn (rare).
• Asked for more context → brief was too thin. Add the missing fact and re-spawn (rare).
• Succeeded → brief was right-sized. This is the common case if you followed the five rules.

Most "subagent failures" are brief failures, not subagent failures.

Red Flags in Your Own Output

If you catch yourself writing any of these in a subagent prompt, stop and re-compress:

• Three or more paragraphs of background before the actual task
• Verbatim file contents longer than 20 lines
• The phrase "for context, here is..."
• Lists of things the subagent "might want to know"
• More than one task per subagent (split into separate Task calls instead)

Why This Skill Exists

Multi-agent serving is a 2026 research topic, not a shipped feature. arXiv 2604.25899 (Pythia, 2026-04) demonstrates that exposing multi-agent topology to the serving layer can recover 3-5× of wasted tokens via prefix sharing — but no production LLM provider has shipped this yet. Anthropic, OpenAI, Codex: every subagent today is a cold-start, full-prefix re-tokenization.

Until that changes, the brief is the only knob you have. Use it.