Dispatch

dispatch — the right worker for the job

A subagent inherits the main loop's model unless you say otherwise. That means a grep across a directory and a database migration both run on Opus, and you pay Opus rates for both. Most delegated work does not need the best model in the lineup.

The output-token spread is 10× end to end: Haiku $5/M, Sonnet $15/M, Opus $25/M, Fable $50/M. That spread pays off on volume, not on any single call — so route every spawn, and never agonize over one.

The lanes

Ask one question: does this agent write or run anything?

A judgment the parent will act on is Build-lane even when nothing gets written — reconciling conflicting sources, picking between options, deciding whether something is sound. Plain extraction and summary stay in Read.

"One batch" means spawned together in a single dispatch. Eight agents run two at a time across a task are eight Read-lane agents, not a swarm.

Opus is the ceiling. Never auto-route to fable; a subagent smarter than the parent means judging output the parent could not have produced. It unlocks only when the user writes the word — fable, or a specific model ID. A capability request ("use the best model", "this one's hard, go all out") is not naming it: that caps at opus.

Haiku exists only in the swarm lane. A lone Haiku agent is never the right call — if it's worth one agent, it's worth Sonnet.

The grok lane — a different meter

grok is not a fourth tier. It is a Grok 4.6 subagent on the Cursor subscription, spawned through Bash, not through the Agent tool. Its work never touches the Claude allowance; only the answer it hands back enters context. Measured on a whole-repo read: 291k tokens spent there, 4.7k returned here.

So route this lane by verifiability, not by strength. Send grok anything token-heavy whose answer you can check cheaply — however hard it is:

• whole-repo or whole-directory reads
• log, output and test-failure sweeps
• broad multi-file searches where only the conclusion comes back
• first-draft implementations that ship with tests you will run

Keep on the Claude lanes anything you cannot cheaply check: design judgment, research verdicts, anything writing to a crux vault, and any call where being subtly wrong costs more than the tokens saved. Grok output is untrusted input — verify before acting.

grok -r -e high "..."          read-only analysis   (the default posture)
grok -e high "..."             write and run enabled
grok -r -b brief.md "..."      splice a written brief in

Effort per task: xhigh for deep analysis where the answer is the deliverable, high by default, medium for mechanical sweeps, low for smoke tests only. Higher effort spends the other meter, so when in doubt go up.

Grok runs in its own process and cannot see this conversation. A thin answer almost always means a thin brief. Mechanics, brief format and failure modes are in references/grok.md — read it before the first grok call of a session.

Quota gate

Before a batch that will spend real tokens, check what is left:

~/.claude/bin/claude-quota

Gate at 80 rather than 95: the turn that checks the quota also spends it. The 7-day window gets the tighter gate because it refills over days, not hours.

Never read a missing cache as "near the limit." NO DATA is the normal state on this machine — the cache is only written when Claude Code invokes the custom statusLine command, and the desktop app does not. Treating absence as a limit signal would route everything to grok forever, silently.

Escalations

Four conditions override the lane and send the work to opus:

1. A lower lane already came back thin, wrong, or hedged. Retry one lane up. Do not re-prompt the same tier with a firmer instruction — the tier was the problem. This escalation stops at opus — it never reaches for fable. When an Opus agent comes back thin, the model was not the bottleneck: rewrite the brief, split it into narrower briefs, or do the work inline. Re-spawning on a sharper brief is fine — re-spawning the same brief is what never works.
2. You can't tell which lane it is. Ambiguity resolves upward.
3. The brief is ambiguous or underspecified. An agent that must infer the goal needs the judgment to infer it well.
4. A read feeds something security- or correctness-critical you won't re-verify. Read-only does not mean low-stakes.

Everything else stays in its lane. Don't invent a fifth condition mid-task.

Depth

Do not set an effort parameter — subagents inherit the session's effort, and that's correct. Shape depth through the prompt instead, which costs nothing:

• Swarm agents: "Answer from the first solid match. Don't verify, don't survey alternatives."
• Build agents on design or root-cause work: "Verify by a second method before returning."
• Read agents: neither — let them work at their natural depth.

This rule is about the Agent tool. grok -e is a different knob on a different meter — set it per task, as the grok lane says.

Say what you routed

Two things, both cheap:

• One line before a batch, before the spawn calls: "7 haiku for the sweep, opus for the implementation." When an escalation fired, name it in the same line: "opus — the sonnet pass came back hedged."
• A model tag in every agent's description, e.g. review:auth [opus], so the routing is visible in the agent list without opening anything. Grok calls get the same tag in the Bash description: sweep crux engine [grok:xhigh].

Mechanics

Per call on the Agent tool — model takes the bare tier name, not a full model ID:

Agent({ description: "map auth flow [sonnet]", model: "sonnet", prompt: "..." })

Inside a Workflow, the same lanes apply to each agent() call:

agent(prompt, { model: 'haiku', label: 'sweep:api [haiku]' })

Omitting model inherits the main loop. That is a routing decision too — make it on purpose, not by forgetting.

What this is not

• Not for the main loop. Dispatch routes subagents. Your own model is the user's choice.
• Not a reason to spawn. If the work is faster done inline, do it inline. Routing a subagent you shouldn't have spawned saves nothing.
• Not a swarm generator. Five-plus agents is a precondition for Haiku, not a target. Don't inflate a fan-out to unlock the cheap tier.

Quick reference

writes / runs / implements / designs / consequential   → opus
judgment you'll act on, even with nothing written      → opus
read-only, single agent                                → sonnet
>5 read-only agents in one batch, breadth sweep        → haiku
unclear · underspecified · read feeding critical
  unverified work                                      → opus
sonnet/haiku came back thin                            → same task, one lane up
opus came back thin                                    → sharpen the brief, same lane
"use the best model" (fable not named)                 → opus

token-heavy AND you can check the answer cheaply       → grok  (Cursor meter)
token-heavy but you cannot check it                    → opus
5h ≥ 80% or 7d ≥ 75%, reading fresh                    → shift checkable work to grok
quota reading stale / missing / NO DATA                → route normally, say so