Blog · August 6, 2026 · 7 min read · by Vilva Athiban P B

Per-Subagent Model Selection: Route the Grunt Work Down, Keep the Judgment Up Top

Claude Code recently gained one of those features that sounds like a footnote and turns out to change how you work: you can now assign a different model to each subagent. Your orchestrating session can run on a strong model while a fleet of helpers — formatters, test runners, doc writers, researchers — runs on cheaper, faster ones.

Done well, this cuts your token bill substantially and speeds up multi-agent runs, because the routine work stops queuing behind expensive reasoning. Done badly, it means your hardest problem gets handed to your weakest model and you spend the savings debugging. This guide is about doing it well: how per-subagent model selection works, a simple mental model for choosing tiers, and configs you can copy.

How per-subagent models work

Subagents live in Markdown files with YAML frontmatter — in .claude/agents/ inside a project, or ~/.claude/agents/ for all projects. (If subagents are new to you, start with our subagents guide and come back.) The frontmatter now accepts a model field alongside name, description, and tools:

---
name: test-runner
description: Runs the test suite and summarizes failures
tools: Bash, Read, Grep
model: haiku
---
Run the relevant tests for the change being discussed.
Report failures with file, line, and a one-line cause.
Do not attempt fixes unless explicitly asked.

You can name a tier like haiku or sonnet, or omit the field entirely — in which case the subagent inherits from the main session. That inheritance is the important default to understand: if you never set anything, every helper you spawn runs on your primary model, at your primary model's price.

A mental model: interns, engineers, architects

The clean way to think about tiers is to ask what happens when the agent is wrong.

Intern work (cheapest tier): if a mistake is instantly visible and trivially re-runnable, use the cheapest model. Formatting, running test suites and reporting output, renaming across files, generating boilerplate, summarizing logs. The task is mechanical; the transcript is the check.
Engineer work (middle tier): if a mistake would survive a casual glance but get caught by review or tests, use a mid-tier model. Writing unit tests, drafting documentation, first-pass code review, dependency-update triage, research summaries.
Architect work (strongest tier): if a mistake would quietly ship — subtle logic, security review, tricky debugging, anything cross-cutting — pay for the strongest model. This includes the orchestrator itself, because routing and synthesis errors poison everything downstream.

Most people discover their agent roster is 70% intern work being billed at architect rates. That is the gap this feature closes.

A starter roster

Three agents cover most workflows. Adapt the prompts to taste:

# .claude/agents/formatter.md      -> model: haiku
# .claude/agents/test-runner.md    -> model: haiku
# .claude/agents/code-reviewer.md  -> model: sonnet (or your main model)

Keep your main session on the strongest model you can afford and let it delegate aggressively. The economics compound: a refactor that spawns five formatting-and-test cycles now runs four of those five cycles at a fraction of the cost, and usually faster — the smaller models respond quicker, so mechanical loops tighten.

The failure modes to avoid

Downgrading the judge. Never give the cheap model the final say on quality. A haiku-tier test runner should report results; the decision about what a failure means belongs upstream.
Cheap research, expensive conclusions. If a weaker model gathers the facts your strong model reasons over, errors launder themselves into confident answers. For research that feeds decisions, spend the tokens.
Forgetting inheritance. An unset model field silently bills at your top rate. Audit your agents once: thirty seconds per file, real money per month.
Tier-splitting a conversation. Subagents don't share context with your main session. That's a feature for parallel work, but it means the cheap agent can't "ask" the smart one for nuance mid-task — so keep intern tasks genuinely self-contained.

The part nobody warns you about

Here's the irony of getting this right: the better your model routing, the less reason you have to watch the terminal. A well-tiered fleet runs longer, cheaper, and more autonomously — which means the finish line moves from "whenever you happen to glance over" to "whenever the run actually ends," and those two moments can be twenty minutes apart. If you've already read our piece on not babysitting the terminal, you know where this is going: the point of cheap, parallel, well-routed subagents is reclaiming your attention, not relocating it.

Start here today

Run /agents, open each existing agent, and add an explicit model line — cheapest tier for anything mechanical.
Create one haiku-tier test-runner or formatter if you have none.
Watch your usage for a week and promote any agent whose mistakes cost you review time — demotions are free, bad output isn't.

Model selection per subagent is the rare optimization that improves cost, speed, and focus at once. Route the grunt work down, keep the judgment up top, and let the fleet run.

Keep reading

Desk Exercises for Programmers: 15 Moves to Do While Your AI Agent Codes

7 min read

Exercise While Coding: Turn Claude Code Wait Time Into Micro-Workouts

7 min read

Claude Code Notch Notifications on Mac: See When Claude Is Done Without Watching the Terminal

7 min read

Claude Code Tips and Tricks: 12 Ways to Use Claude Code Effectively

8 min read

Claude Code Multiple Sessions: How to Run Agents in Parallel Without Losing Track

6 min read

A Claude Code Workflow That Doesn't Involve Watching the Terminal

5 min read

Claude Code Hooks: A Practical Guide to Automating Your Agent Workflow

7 min read

Claude Code Auto Mode: Fewer Permission Prompts Without Living Dangerously

7 min read

Claude Code Subagents: How to Delegate Work to Specialized Agents

7 min read

Claude Code Context Management: Treat the Context Window Like a Budget

6 min read

Skills and Plugins: How to Teach Claude Code Your Way of Working

7 min read

Claude Code Background Tasks: Run Long Commands Without Blocking Your Session

6 min read

Codex CLI Notifications: How to Get a Ding When Codex Is Done or Needs Input

7 min read

Cursor Notification When Done: Every Way to Get Notified When Cursor Finishes

6 min read

Claude Code Notifications: How to Get Notified When Claude Code Finishes or Needs Your Input

6 min read

Want to Be Notified When Claude Responds? How Claude Notifications Work on Web, Desktop, and Mobile

5 min read

Claude Code Notification Scripts: Copy-Paste Recipes for Every Platform

6 min read

Get a Ding the Moment Codex Needs Your Response

6 min read

Get Notified the Moment Claude Code Is Waiting for Your Input

6 min read

“Notifications Are Turned Off for Claude” — Here Is the Fix

6 min read

Codex Sound When Done: Make Codex CLI Play a Sound When It Finishes

6 min read

terminal-notifier + Claude Code: Native macOS Alerts When Your Agent Finishes

6 min read

Gemini CLI Notifications: How to Get a Sound or Alert When Gemini Finishes

6 min read

Get Claude Code Notifications on Your Phone

6 min read

Claude Code Remote Control, Explained

7 min read

preferredNotifChannel: Claude Code's Built-In Notification Setting

6 min read

Claude Code Notifications in tmux and Over SSH

7 min read

Claude Code Effort Levels: Why Your Setting Keeps Getting Ignored

8 min read

Codex vs Claude Code Notifications: How Each One Tells You It Is Done

8 min read

Claude Code Notifications Not Working: A Diagnostic Checklist

10 min read

Claude Code Notifications Inside Your Editor's Terminal

7 min read

CLAUDE_CODE_DISABLE_BACKGROUND_TASKS Explained

7 min read

CLAUDE_CODE_DISABLE_BG_SHELL_PRESSURE_REAP Explained

8 min read

Cursor Alerts Explained: Every Alert Cursor Raises, and How to Control Them

7 min read

Claude Code Environment Variables: The Groups That Actually Matter

10 min read

Claude Code Notifications on WSL and Windows

9 min read

Claude Code Checkpoints and /rewind

8 min read

Claude Remote Control: How It Works, Turning It Off, and Auto Mode

7 min read

Claude Code /loop: Poll a Deploy, Babysit a PR, and Get Told When It's Done

8 min read

Setting Up NotchFit With OpenAI Codex CLI: Hooks, $notchfit-plan, and the Daily Flow

7 min read

Claude Mods Explained: What the New Plugin Layer Can Do, and How to Turn On 'You Should Know'

9 min read

NotchFit vs Mac Break Reminder Apps: A Clock or a Signal From Your Coding Agent?

8 min read

Claude Code Headless Mode: Running claude -p in Scripts and CI, and the Three Signals That It Finished

10 min read

Building a Movement Habit at a Desk: Use the Agent Run as the Cue, One Set as the Unit, and Streaks Carefully

8 min read

Claude Code MCP Servers Explained: Adding Servers, Choosing a Scope, Handling Secrets, and Why Long Tool Calls Go to the Background

10 min read

/notchfit:plan Explained: How Claude Code or Codex Writes a Weekly Workout Plan the Notch Can Actually Run

8 min read

Claude Code Spinner Verbs Explained: What “Pondering” and “Brewing” Mean, How to Change Them, and Why You Should Stop Watching the Spinner

7 min read

Why NotchFit's Done and Permission Alerts Wait Until Your Set Ends, When That Costs You, and How to Turn the Workout Off

7 min read

How to Resume a Claude Code Session: --continue vs --resume, Naming Sessions, the Session Picker, Branching, and What Comes Back

9 min read

NotchFit on a Mac Without a Notch: How Pill Mode Works on Mac mini, iMac, Mac Studio and a MacBook With the Lid Closed

6 min read

How to Run a Claude Code Session in the Background: /background, claude --bg, Agent View, and Getting Back to It

9 min read

What a Notch App Can See: NotchFit's Local-First Design, What the Claude Code and Codex Hooks Send, and What Stays on Your Mac

7 min read

Why Codex CLI Asks for Permission Every Time, and How to Tune the Sandbox and Approval Policy Without Going Full Access

8 min read

What to Do While Claude Code or Codex Works: A Simple Protocol for the Two-to-Twenty-Minute Gap in Every Agent Run

7 min read

Is It Possible to Get a Notification When Claude Code or Codex Hits a Reset? Yes for Claude Code: the quota_auto_resume Hooks, Explained

8 min read

What a MacBook Notch App Can Actually Do: The Five Kinds of Notch Utility, Their Trade-offs, and Where NotchFit Fits

7 min read

How to Make MCP Servers Not Take Up as Much Context in Claude Code: Tool Search, alwaysLoad, Scopes and Output Caps

8 min read

The notchfit Command Line: Drive Your Workout, Read Status as JSON, Validate Plans and Test the Notch Without Leaving the Terminal

7 min read

How to Audit MCP Tool Calls Made by Claude Code: A Five-Minute Hook Audit Log, OpenTelemetry Tool Events, and the Transcript for Investigations

9 min read

Exercise Snacks: What the Research Actually Says About One-Minute Vigorous Breaks, What It Does Not, and How NotchFit's Agent-Run Sets Compare

8 min read

How Long Until a Claude Code MCP Tool Call Times Out? The Four Timers (Startup, Two-Minute Backgrounding, Idle, Wall-Clock), Their Defaults, and Which One to Change

8 min read

Sitting Too Long as a Developer: The Posture Problems Coding Agents Made Worse, What the Research Says About Breaking Up Sitting, and How NotchFit Turns an Agent Run Into the Break

8 min read

Claude Code .mcp.json Explained: Where Project-Scoped MCP Servers Are Read From, What Happens in Subfolders and Monorepos, and Why a Committed Server Shows Pending Approval

8 min read

Standing Desk Exercises for Developers: Why Standing Still Is Not the Goal, Eight Movements That Fit a Thirty-Second Agent Run, and How NotchFit Cues Them From the Notch

8 min read

The claude mcp Command Reference: add Syntax by Transport, What --scope local, project and user Actually Mean, add-json and add-from-claude-desktop, list, get, remove, login and serve

9 min read

Your First Week With NotchFit: Install and Simulate, Run Real Tasks, Write a Smaller Plan Than You Want, Tune the Alerts, Survive the First Skipped Day, and Decide What Stays

8 min read

How to Stop a Running Claude Code Task: Esc vs Ctrl+C, Background Tasks, Subagents, and Disconnecting Remote Control

8 min read

Pomodoro vs Agent-Driven Breaks: Timing Movement to Claude Code and Codex Runs Instead of a Timer

7 min read

Claude Desktop MCP Configuration: claude_desktop_config.json, Logs, Extensions, and the Difference From Claude Code

9 min read

The 20-20-20 Rule for Developers: Eye Strain, the Evidence, and Agent Runs as the Cue

8 min read

Claude Agent SDK + MCP Servers: mcpServers, allowedTools, Status Checks and Auth

9 min read

Discreet Desk Exercises for an Open-Plan Office While Claude Code Runs

7 min read

claude remote-control Server Mode: Run Claude Code as a Session Server

9 min read

Wrist and Forearm Stretches for Developers: A 60-Second Routine for Every Claude Code Run

7 min read

Claude Remote Control vs Claude Code on the Web: Which to Use and How They Differ

9 min read

Neck and Shoulder Stretches for Developers: A 90-Second Routine Between Agent Runs

7 min read

What Happens When You Step Away From a Claude Code Session: Prompt Cache Expiry, Idle Compaction, and When to Come Back

8 min read

Hip Flexor and Lower Back Stretches for Developers: A Two-Minute Standing Routine Between Agent Runs

7 min read