Is this skill against using subagents?

No. It argues against dispatching to save money on short turns. Independent, parallel or context-heavy work such as exploration, adversarial review or long refactors is where subagents pay off, and the skill tells your agent to keep working while they run.

Where do the numbers come from?

From an 18-day experiment with ModelSelector, a prompt-scoring auto-router used in daily Claude Code work. The router was retired after the data showed the cold-start cost. The full postmortem is public and linked in the skill.

Does it work outside Claude Code?

The decision rule applies to any agent that can spawn subagents, such as Codex CLI. The examples use Claude Code's Agent tool and hook names.

I'm building a model router. Will this help?

Yes. It explains why complexity-score routing and skip lists fail, and gives the positive-trigger whitelist to use if you still need a router.

One SKILL.md with the cost model, the four-condition decision rule, three anti-patterns, verification signals and worked examples. No scripts or dependencies.

Subagent Dispatch: When to Delegate vs Just Answer

Example session with this skill installed

I'm adding a UserPromptSubmit hook that sends simple prompts to Haiku to save tokens. Most of my prompts are short questions and follow-ups. Good idea?

Read your context and instructions
Compiled the subagent dispatch

Short answer: no. For short questions and follow-ups, a dispatched subagent costs more than the direct reply it replaces.

Why

Path	Tokens per turn	What you pay for
Direct reply from the main model	1-3k	Cache-warm input plus a short answer
Dispatch to a smaller model	30-60k	Context repacked into the subagent prompt, the subagent's work, the result relayed back

A smaller model is about 5x cheaper per token, but the cold start costs 5-20x more tokens, so the hook loses on exactly the prompts it
was built for. Cached input on the main model is already billed at a fraction of the normal rate, which is why the direct reply is so
cheap.

What to build instead

Default to a direct reply. Dispatch only when all four hold: the task is independent, the output is substantial (over 10k tokens),
nobody will ask follow-ups mid-task, and the handoff fits in a self-contained prompt under 2k tokens.
If you still want a hook, trigger on positive signals only: explicit file paths, build, test and git commands, words like "refactor",
"migrate" or "across all files".
Never route by a complexity score, and never keep a skip list of conversational phrases. Both fail in ways that only show up in the
token bill.
Leave "use a cheaper model" to an explicit choice by the user.

How to check

Compare tokens per session with and without the hook for a week. If a two-sentence question ever costs more than 40k tokens, the hook is
dispatching conversation.

Connects securely to your tools. The creator never sees your data.

About this skill

Handing a short question to a "cheaper" subagent feels like saving money. Measured, it usually is not: the subagent starts cold, the context has to be repacked into its prompt, and the result has to be relayed back. In an 18-day experiment with a prompt-scoring auto-router in daily Claude Code use, that cold start cost 5-20x the tokens of a direct reply on short turns, and the router was retired. This skill gives your agent the rule that came out of it.

What your agent does with it:

Dispatches a subagent only when four conditions all hold: an independent task, substantial output, no multi-turn dependency, and a self-contained handoff prompt.
Answers directly for questions, recall, clarifications and anything multi-turn.
Keeps working while dispatched agents run instead of waiting on them.
Recognises three anti-patterns: complexity-score auto-routing, a weak local LLM preprocessing context for a cloud model, and skip-list blacklists.
Checks afterwards whether the call was right, with concrete signals for both outcomes.

Also useful when you are designing a UserPromptSubmit hook, a model router, or a CLAUDE.md rule that forces delegation.

What's inside: one SKILL.md with the cost model, the decision rule, anti-patterns, verification signals and worked examples, plus a link to the public postmortem of the experiment.

Compatibility: written for Claude Code's Agent tool. The decision rule applies to any agent that can spawn subagents. No tools, scripts or MCP servers required.

How to use

Skills on Agensi run inside your AI agent. Copy the prompt, paste it into your agent, and it connects to Agensi and installs this skill for you. If a payment is needed, your agent shows it to you first.

Reviews

No reviews yet

Be one of the first to try it. Every listed skill passes our trust checks below.

Security scanned

Passed our 8-point scan before listing

Fresh listing

Recently published to Agensi

Money-back guarantee

Doesn't do what its page promises? You get your money back.

Subagent Dispatch: When to Delegate vs Just Answer