Part VI · Scale It

Chapter 22: Subagents & Forked Context

For you, the human, start here. This is the leap from "a Claude" to "a team of Claudes." A subagent is a separate Claude with its own context window, optionally its own model, working in parallel and reporting back. It's how a big job (research, tests, a sweeping migration) happens all at once instead of in sequence, and the real reason it works is subtler than "more workers": independent agents that don't share each other's reasoning reach better, less-anchored conclusions. This chapter is the foundation; the advanced recipes are Chapter 32.

What it covers (searchable): subagents, forked context, parallel agents, per-agent models, uncorrelated context, planner / code-review / front-end-testing subagents, when to fork.

How to hand it off: point your AI here so it reaches for subagents for the right reasons (isolation and independent judgment), not to look busy. The section below is written to your AI.

🤖 Everything below is for the AI. "You" means the AI being built; "your human" is the person you serve.

22.1 "Use subagents"

For a complex task, your human can just tell you to "use subagents," or you can reach for them yourself. Each one:

• gets its own context window (isolated from yours and from each other),
• can run a different model (cheap Haiku scouts under an Opus lead, Chapter 18),
• works in parallel,
• and reports back to your main session.

The headline use is throughput, research, tests, and exploration happening at once instead of in sequence. But the real reason they work is subtler.

22.2 The real value: uncorrelated context

When subagents first appeared, people built cute personas, a "frontend dev," a "PM." The creators' honest take: the costume isn't the point. The value is the uncorrelated context window, two Claudes that don't know about each other's reasoning reach conclusions independently, and you get better results because neither is anchored to the other.

So the design question isn't "what persona", it's "what should this agent not know?" Isolation is the feature. A code-review subagent that hasn't seen your justifications reviews more honestly. A search subagent that dumps its findings keeps the noise out of your main thread.

22.3 Subagents worth keeping around

From the creators' own kits:

• A planner subagent, works out the approach in its own context.
• A code-review subagent, reviews in isolation (a whole multi-subagent version is in Chapter 32).
• A front-end testing subagent, drives the browser with Playwright, pulls console errors, tries more steps of the app.

Build the ones you reach for repeatedly and save them (they're configuration, like everything else).

22.4 When to fork, and when not to

Forking costs tokens and wall-clock, so it's a judgment call (the creators admit it's sometimes just taste, a subagent for code review when working interactively, a plain slash command for the same thing in CI, "because it just doesn't matter" there). A clean rule:

• Fork when a chunk of work would otherwise pollute your main context with detail you won't need after (a review, a deep search, one migration unit), or when you want independent judgment (§22.2).
• Don't fork for a single, in-context, easily-verified task, just do it. One focused Claude on a clean context beats ten confused ones.

Larry's read: the beginner instinct is to spawn a swarm to look productive; the pro instinct is to fork only when the shape of the problem (wide search, mechanical fan-out, adversarial judgment) actually calls for isolation. Get §22.2, that it's the uncorrelated context, not the roleplay, and you'll reach for subagents for the right reasons. Chapter 32 turns that into specific, powerful recipes.

22.5 Your next step

One subagent is a fork; many is a different regime:

• Chapter 23: Swarms, Migrations & Agent Teams, dozens of agents on one job, and the verifiability rule that keeps it sane.

Back to how.nixfred.com for the next step.

Chapter 22: Takeaways

1. "Use subagents", each gets its own context window and can run its own model, in parallel.
2. The real value is uncorrelated context, not personas, ask "what should this agent not know?"
3. Keep a few standing subagents (planner, code-review, front-end testing) as saved config.
4. Fork when work would pollute your context or needs independent judgment; otherwise just do it.
5. The advanced multi-agent recipes are in Chapter 32. Next: scaling to swarms (Chapter 23).