Claude Code subagents aren't multithreading — the hidden cost of "use subagents liberally"

There's a CLAUDE.md file going viral right now that includes the instruction: "Use subagents liberally." It sounds like a power user tip — but it's expensive advice that misunderstands how subagents actually work.

The mental model people have

If you've worked with multithreading or parallel processing, subagents feel familiar. Spin up more workers, get results faster. More parallelism = more speed.

But subagents aren't threads — and most of the time, it's better to let Claude decide when to use them.

What actually happens when a subagent spins up

Every subagent starts cold. Before it does a single useful thing, it needs to:

• Bootstrap its context (~20,000 tokens)
• Load tool definitions and verify access
• Process the task description you gave it

That's the equivalent of booting a new container for every task. A grep that takes <1 second inline costs 20k+ tokens and several seconds through a subagent.

When subagents actually help

Subagents aren't bad — they have a specific sweet spot:

• Parallel, long-running tasks — the startup cost gets amortized across many operations.
• Noisy output isolation — verbose results stay in the subagent's context, only a summary comes back.
• Deep exploratory research — 10+ tool calls to find what you need, better to keep that out of your main context.

When they hurt

For anything that's 1-3 tool calls — a quick search, reading a file, a targeted edit — inline is both faster and cheaper. And that's most tasks.

Telling Claude to "use subagents liberally" means it reaches for the expensive option on tasks where the overhead dwarfs the actual work. Multi-agent sessions consume 3-4x more tokens than single-threaded operations.

Power users are already filing bugs

Search the Claude Code GitHub issues for subagent complaints. Every single one is about Claude spawning too many subagents, not too few:

• Subagent spawning wastes tokens on tasks that should be done directly — subagents used for simple find-and-replace work
• Subagents too slow and cost more tokens — exactly what it says

Blog posts tell the same story: one developer burned 887k tokens/minute with aggressive subagent use. DEV Community posts titled "Claude Code Sub Agents - Burn Out Your Tokens" and "Why Each Subprocess Burns 50K Tokens".

Not a single issue asking Claude to spawn more subagents. The complaints all go one direction.

The thing is, Claude already knows the difference

Claude Code already weighs this tradeoff natively. It reaches for subagents when isolation or parallelism genuinely helps. When you override that judgment with "use subagents liberally," you're not unlocking a superpower — you're just turning off the part of Claude that was saving you money.

Beyond subagents

This pattern applies beyond subagents. A lot of CLAUDE.md files I see include instructions that restate behavior Claude already has, or nudge it toward patterns that sound productive but cost you real money.

The most effective CLAUDE.md files focus on what Claude can't infer on its own — your project's non-obvious gotchas, the things that look right but break in production, the hard-won lessons from debugging sessions. Not generic workflow advice.