Putting model: haiku on a scout is the cheapest optimisation in agent work: a search-and-summarise job rarely needs a frontier model. It only works if the host actually runs that model. This chapter shows how to check, and what to do when it does not.
How hosts resolve model:
| Value | Claude Code | Cursor |
|---|---|---|
haiku, sonnet, opus | Latest model of that family | Cursor's own alias mapping, which may differ from Claude Code's |
Full ID like claude-… | That exact model | That model, if enabled in your settings |
inherit or missing | Same as the main chat | Same as the main chat |
The alias difference matters. When last checked for the Claude Code Subagents for Cursor README, Cursor mapped haiku to an older Haiku that many plans have disabled, while Claude Code mapped it to its latest Haiku. Same file, different model.
The two cases where Cursor ignores your agent's model
Cursor's Task tool runs the subagent on the parent chat's model when:
- You use a personal Anthropic API key (BYOK). Cursor staff confirmed this on the Cursor forum.
- The model named in the agent file is not enabled in your Cursor model settings.
Nothing in the reply tells you. You find out from the usage meter.
Check which model ran
Before you optimise, measure.
- Give the scout a trivial, recognisable task: "List the top-level folders."
- Run it once with your parent chat on an expensive model.
- Check your provider's usage dashboard or request logs for the model of that request. In Claude Code, the transcript and
/costshow usage for the session. - If the request was billed to the parent's model, you have hit one of the two cases.
Fix 1: enable the model (Cursor's own models)
If you are not using BYOK, open Cursor's model settings and enable the model your agent names. Then re-run the check. This is the only change needed for many setups.
Fix 2: route subagents through Claude Code
If you use BYOK, or you want Claude Code's alias mapping, run the subagent through Claude Code's headless mode from Cursor's shell. Claude Code Subagents for Cursor is a skill that does this routing:
npx skills add ysskrishna/claude-code-subagents-for-cursor -a cursor -gThen, in a Cursor chat:
/use-claude-code-subagentsThe agent confirms Subagent mode: claude code. Every custom subagent now runs through claude -p on the model in its own file, and each answer ends with a receipt like:
scout ran on claude-haiku-…Requirements from the README: Cursor desktop running locally (Cloud Agents have no claude CLI), Claude Code 2.1.281 or later signed in, and a POSIX shell (macOS or Linux; native Windows is not supported). Runs bill your Claude Code account. Say "use cursor subagents" to switch back.
Permissions when routing
Claude Code runs outside Cursor's sandbox, so pick a Cursor run mode that allows it (Auto-Review with Sandbox is the suggested one; Allowlist modes are not supported). On the Claude Code side the skill passes --permission-prompts none, so anything Claude Code would ask about is denied and reported, and only editing agents get --permission-mode acceptEdits.
Choosing models by job
| Job | Model tier | Why |
|---|---|---|
| Repo search, file listing | Small (Haiku class) | Mostly reading; output is a list |
| Log or PR summarisation | Small | Compression, not reasoning |
| Code review | Mid (Sonnet class) | Needs judgment, not maximum depth |
| Architecture, tricky debugging | Large, usually the main chat | Keep it where the context is |
Treat this as a starting point and measure. A subagent that misses things on a small model costs more in re-runs than it saved.
Quick check
-
You use BYOK in Cursor and your agent says
model: haiku. What model runs via the Task tool?The parent chat's model. -
What is the receipt line for, beyond reassurance?It catches routing mistakes per run, before they show up on the bill.
-
Why might
model: haikurun a newer model through Claude Code than through Cursor?The two hosts resolve the alias differently; Claude Code maps it to its latest Haiku.






