The helper gets a clean window

Lesson 0010 still fills your RAM if you do every grep. A sub-agent is a second loop with its own window. You keep the plan. They keep the dumps. You take back a short summary.

What this is. Anthropic’s third long-horizon lever, after compaction and notes: specialized helpers with clean context windows. Each may spend tens of thousands of tokens exploring; the lead gets a distilled summary (often 1–2k). Effective context engineering · “Sub-agent architectures”

What this is not. A second model (it often inherits yours). A skill (that plays out in your thread). Agent teams, worktrees, or a 15-agent research swarm — parked. Not a ChatGPT menu item named “Subagent.”

1 · IsolateHelper RAM is a second bar. Search dumps never hit the lead.
2 · BriefObjective, format, bounds. Vague briefs duplicate work.
3 · ReturnA short note, not the transcript. Then the lead picks.

Same agent loop as 0006. Same harness as 0007. Tonight is whose window the greps land in.

Two context-window bars. The lead bar is thin: plan plus a small summary card. The helper bar is packed with grep and file dumps. A dashed arrow returns only the summary.

The lead coordinates. The helper explores.

One agent trying to hold the whole project is how windows die. The lead keeps a high-level plan; helpers do the noisy search. Separation of concerns: detailed search stays isolated; the lead synthesizes. The research-system writeup names the same move as compression — distill a corpus in a window you are willing to throw away. How we built our multi-agent research system

Leadplan + brief

Helperown RAM, greps

Returnsummary only

Picklead decides

You are sitting in this: Cursor’s Explore / Bash / Browser helpers exist because those jobs spew logs. Cursor · Subagents

Left: one window filling with file dumps, labeled lead does every grep. Right: two windows, lead stays thin, helper is full, arrow says summary only.

One ticket · last hour’s stone

# SOLO JIT — lead does every read
Grep("temp_c", "oven/") → 8 log hits
Read(p1) Read(p2) … Read(p8)     # 6k+ tokens now in LEAD RAM

# SUB-AGENT — second window
Agent({
  type: "explore",
  prompt: """
    Objective: last eight oven probes.
    Format: three lines (start, end, target).
    Bounds: oven/ only. Return no dumps.
  """
})
# helper greps in ITS window
# lead receives:
#   180°C → 188°C. Target 240. Unresolved: preheat.

What this shows. The question did not change. Whose RAM paid for the logs did. Cursor: the parent must put the needed facts in the brief — the helper has no prior conversation. Anthropic: teach the orchestrator to delegate (objective, format, tools, bounds) or helpers duplicate searches.

Head chef at a pizza pass holds a short brief. Through a swinging door, a prep cook is buried in oven-log printouts and sends one small slip back.

The brief is the match

Same lesson as a skill description (0008). Claude Code uses the sub-agent’s description to decide when to delegate. A prompt that says “look into the oven” is how you get refunds/ and a 2021 stone story in the summary. Write what success looks like, what to ignore, and how short the return should be. Create custom subagents

Skill: watch the lead’s bar

Solo is JIT in the lead window. Isolate opens a helper bar. Vague brief still pollutes the return. Sharp brief is the three-line note.

Same ticket, two windows

Isolatesecond RAM

Briefobjective · bounds

Returnsummary only

Lead window

Helper window

Skill in the thread, helper off-stage

Skill (0008)

Chapter loads into your window. You see and steer each step. Use when the procedure is the work.

Sub-agent (tonight)

Side task that would flood you with logs. Isolation is the feature. Cursor: if it is a one-shot format job, it is a skill.

You already sit in one: Cursor’s Explore is a read-only helper so codebase search does not bloat the parent. Claude Code’s Explore is the same idea. The product names differ; the rent is the second bar.

Parked: spinning many helpers in parallel (Anthropic’s Research feature; ~15× the tokens of a chat — only when the task pays). Agent teams, git worktrees, cheaper-model routing. Loop engineering — designing the cycle that prompts the agent — is next, not a second helper.

Check

Primary source. The sub-agent paragraph in Effective context engineering, then the isolation numbers in How we built our multi-agent research system (search is compression; brief the helper). Product form: Claude Code sub-agents and Cursor subagents — you are sitting in the second one.

Card: sub-agents. Last path: just-in-time. Last folder: skills.