Cowork
Claude Subagents: When They Save vs Double Tokens
Claude subagents are parallel or delegated workers. Used well, they cut wall-clock time and keep exploration out of one giant thread. Used poorly, each worker reloads context, tools, and retries on the same ambiguous brief.
This page is part of the Cowork hub, which covers the whole topic end to end.
Not affiliated with Anthropic. Claude Token Optimization is an independent site and audit service. Claude® and related product names are trademarks of Anthropic PBC. Plan prices, model rates, and usage limits change; always confirm on Anthropic’s own docs. [VERIFY: 2026-09-17]
Claude subagents let one parent task fan out. Used well, they avoid stuffing every exploration into one giant thread. Used poorly, they multiply tokens.
Product details evolve, so treat capability claims as items to confirm in current Anthropic documentation. [VERIFY: 2026-09-17] Related: Cowork vs Claude Code and the optimize hub.
What subagents are: the operator view
Think in roles, not magic:
| Role | Job |
|---|---|
| Parent / orchestrator | Holds the goal, splits work, merges results |
| Subagent | Owns a bounded subtask with its own context window and tools |
| Shared pool | On consumer plans, all of this still draws plan usage across session and weekly-style limits [VERIFY: 2026-09-17] |
Subagents help when subtasks are independent, clearly scoped, and mergeable. They hurt when every worker needs the full history or the brief is still fuzzy. See the token tax of tool use and agents for the rounds × tokens model.
The cost lever: save vs double
When subagents save tokens
- Narrow slices: each worker gets only the files, URLs, or document sections it needs.
- Clear Definition of Done: “return three bullet findings plus sources,” not “look into this.”
- No overlapping research: partitions do not redo the same search.
- Cheaper model for scouts, heavier model for merge: baseline examples are Haiku 4.5 $1/$5, Sonnet 5 $2/$10, Opus 5 $5/$25, and Fable 5.1 $10/$50 per MTok input/output. [VERIFY: 2026-09-17]
- Hard caps: maximum steps, tool calls, and wall time per worker.
In that shape, parallelism replaces one mega-thread that re-sends ever-growing history. You pay for several small contexts instead of one gigantic one.
When subagents double or worsen spend
- An ambiguous parent brief is copied to every worker.
- Each subagent loads the same large repository dump or PDF.
- Tool or computer-use loops are unbounded per worker.
- Assignments overlap, such as asking everyone to research the same competitors.
- The parent keeps chatting while workers thrash, then discards their output.
- Fan-out substitutes more agents for a sharper question.
On Pro or Max, that waste appears as faster Settings > Usage burn across the shared pool. With credits enabled, it appears as API-rate spend. [VERIFY: 2026-09-17]
Decision checklist before you fan out
Ask these five questions. If you fail two or more, do not parallelize yet.
- Can I write a one-sentence success test for each worker?
- Can I give each worker a different input slice?
- Do I know how I will merge results without another huge synthesis loop?
- Is wall-clock time worth more than token burn for this task?
- Do I have a kill switch, step cap, or budget per worker?
If the answer to the first or second question is no, clarify the brief in one short thread first. That is classic optimization work.
Patterns that usually work
Map-reduce research. The parent defines questions A, B, and C. Each subagent owns one question and source list. The parent merges them into a decision memo.
Test matrix. Each worker runs one environment or case. The parent compares failures.
Code survey. Each worker owns one directory with a “report only, no edits” rule. The parent plans the patch series afterward, often for Claude Code. See Cowork vs Claude Code.
Scout then snipe. Haiku-class or lighter workers gather candidates; one stronger model writes the final artifact. [VERIFY: 2026-09-17]
Patterns that usually fail
- “Five agents, same prompt, pick the best” without diverse inputs.
- Parallel refactors on the same files.
- Workers told to keep going until confident with no metric.
- Screenshot-heavy UI loops in every worker at once.
These patterns feel productive and empty the weekly bar. Confirm bars in Settings > Usage, then log surface and size with the usage tracker. [VERIFY: 2026-09-17]
Budgeting: plan limits vs API rates
Consumer plans. Pro is about $20/month or $17 annual-equivalent. Max starts around $100/month, with higher per-session depth framed as 5× and 20× relative to Pro. Included usage is not a fixed message count. Subagents accelerate burn of the same pool. [VERIFY: 2026-09-17]
Credits and API. Extra usage credits on paid plans are generally at API rates. Parallel workers make those rates feel sudden. Run volume through the cost calculator before making fan-out a habit. [VERIFY: 2026-09-17]
If you keep hitting ceilings, do not only buy more. Use increase limits as the decision fork: optimize, use credits, or upgrade.
High-fit next step
If parallel workers are chewing through context without clean merges, a written audit of the parent prompts and one worker transcript is higher leverage than another seat.
Have your own prompts looked at →
See the audit example, audit privacy, operator background, and structural playbooks.
FAQ
Are subagents free on Pro if I already pay?
There is no separate subagent fee, but they consume the same usage pool and credits if enabled. Parallelism is not free capacity. [VERIFY: 2026-09-17]
How many subagents should I run?
Start with two well-scoped workers. Add more only when merge quality stays high and each has a distinct slice.
Do subagents help Claude Code too?
Delegation and parallel tool patterns appear across agentic coding surfaces; naming differs by product. Apply the same save-vs-double tests. [VERIFY: 2026-09-17]