choosing-subagent-models
Skill0 starsChanged 50 days agoMITView on GitHub
Use when about to launch any subagent — a research or exploration agent, a reviewer, an implementer, a parallel fan-out, or a fork.
In this repository20 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- templates/CLAUDE.mdCLAUDE.md
- capturing-decisionsSkill
- committing-changesSkill
- debugging-systematicallySkill
- delegating-workSkill
- designing-changesSkill
- exploring-a-codebaseSkill
- injecting-dependenciesSkill
- planning-changesSkill
- pruning-testsSkill
- reflecting-on-changesSkill
- resolving-conflictsSkill
- reviewing-codeSkill
- verifying-changesSkill
- writing-commentsSkill
- writing-logsSkill
- writing-project-instructionsSkill
- writing-skillsSkill
- writing-testsSkill
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 13
Use the aliases haiku, sonnet, opus, not version IDs.
- Distinctive: line 30
Sonnet is the floor for anything that decides correctness: every reviewer, every implementer that must make a test pass.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Choosing subagent models
Every spawn names its model. An omitted model inherits the session's: a lookup from an Opus
session runs at full price, a final review from a cheap session runs cheap, and neither says so.
- The kit's spawn hook refuses a launch with no
model. It never picks one — pick here, by the job, not by the model the session happens to run. - Use the aliases
haiku,sonnet,opus, not version IDs. They survive a model release.
The table
| The subagent's job | Model |
| Mechanical and fully specified — read a known file and extract a value, run a known test, build or lint command and report its output, summarize a log, apply a change whose exact text is given — where one command proves it done | haiku |
| Gathering — web research, codebase exploration, "where is X", reading code to answer a question, implementing a plan task whose code is written out, reviewing one task's diff | sonnet |
| Judgment — design and architecture, debugging after a fix has already failed, security review, the final review of a branch, implementing from prose where decisions are still open | opus |
Where the lines fall
- Haiku gets no search. Unclear where to look, or what the thing is called, makes it research —
sonnet. A file haiku fails to find comes back as a confidently incomplete answer. - Haiku stops at its first failure — a failed command, an edit that does not apply, unexpected
output. Re-spawn on
sonnet; do not retry. Cheap models take 2–3× the turns on multi-step work and cost more overall. A red run fromvvkit:test-runneris its result, not its failure. - Sonnet is the floor for anything that decides correctness: every reviewer, every implementer that must make a test pass.
- Research fleets stay on
sonnet, even for an architectural question. Readers gather; the judgment happens in the main conversation. - Between two tiers, take the higher and say so. When the higher one is a real cost — a fleet, a long run — ask once, naming both tiers and what the higher one buys.
- Escalate, do not repeat. A blocked implementer or a fix loop that is not converging goes up a tier. The same spawn on the same model gives the same result.
- Two failed rounds end re-dispatching. Take the core back into the main conversation, or send a narrower piece one tier up. A wide brief that failed twice fails on scope, and a stronger model inherits the same ambiguity.
Special cases
- A fork runs on the parent's model and ignores
model. Fork only when the inherited context is worth the session's price. - A named agent whose definition sets
modelcarries its author's default for an ordinary job. Pass nothing when the table agrees; pass a tier up when the job is bigger — a whole branch, security-sensitive code. Never a lower one. - The user named a model. Theirs wins.
CLAUDE_CODE_SUBAGENT_MODELforces one model onto every subagent. A debugging tool, not a policy.KIT_SUBAGENT_MODEL_GATE=0turns the spawn hook off. The table still applies.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.