When your main agent hands a browser test to a subagent, the worker you choose affects how clear the task stays. The context is everything the model can see while it answers you. A subagent runs in its own context window, keeping its instructions outside your working view. The catalogue now covers 35,824 subagents, so each definition is worth checking.
Choose the handoff by its job
The main agent chooses a subagent from its description. That description is the whole interface. Choose a narrow role, such as browser tests or review triage, when you want a predictable handoff. The catalogue places 13,212 entries under Claude Code, so a vague label leaves too much to chance.
A subagent’s instructions sit in its own definition file. Each page names the repository file it read and states whether it is the first file or a later copy. The original is the first copy of a file we saw; everything matching it afterwards is a copy. Demonstrate is read from .github/agents/demonstrate.md, and its page says so.
The 43-token price of delegated work
A token is the unit your AI plan bills by, about three quarters of a word. The token figure is the subagent’s name and description, which sit in your main agent’s context so it can pick one. The body of its instructions runs in the subagent’s own window.
The median definition loaded every session costs 43 tokens. One definition in ten costs 109 tokens or more. At 30 sessions per day, the 43-token median becomes 1,290 across a day. The maximum is beam-sre, at 986 tokens every session across 997 lines. That is about the length of a long email placed before work begins.
What the subagent scan sees in its instructions
The scan graded 34,555 entries A, 782 B, 486 C or D, and 1 E or F. Grade A means it found nothing it treats as risky. It does not mean the page lists no finding at all.
The most common findings are making network calls, 1,263 times; recursive force deletion, 254; running shell commands, 225; and reading agent configuration directories, 213. rust-mcp-expert is graded C for recursive force deletion. conversation-analyzer is graded D for asking for root and recursive force deletion. reproduce-verifier is graded C for downloading and executing remote code and making network calls.
Every scan finding is printed with the line that produced it. Read that line as evidence about the instructions, then decide whether the worker suits your repository.
Six focused workers to try first
The tail after each name gives its cost, line count, and copies.
- Demonstrate: tests a pull request’s user interface and records the results, original, 10 tokens every session, 131 lines, 1 copy.
- playwright-test-generator: creates one automated browser test from a supplied scenario, original, 151 tokens every session, 59 lines, 2 copies.
- AVM Owner Triage: sorts Azure Verified Modules issues into delegatable and human piles, original, 61 tokens every session, 683 lines, 2 copies.
- review-triager: sorts pull-request review threads into an action plan and can manage those threads. Original, 0 tokens every session, 132 lines, 1 copy.
- flutter-reviewer: reviews Flutter and Dart code for style, performance, accessibility, and architecture, original, 50 tokens every session, 253 lines, 7 copies.
- code-reviewer: reviews Python, notebooks, scripts, and automation against project standards, original, 52 tokens every session, 206 lines, 3 copies.
When delegation is the wrong shape
A skill is a folder of instructions the model opens when it decides it needs them. It fits better when reusable guidance should appear only for a specialised task. A hook is a shell command your agent runs for you when something happens. It fits better when an action must follow an event.
A subagent’s instructions never sit in your main context, and it is the main agent that decides to call it. Use the agents catalogue when you need a delegated worker, not a conditional rule or automatic action.
Questions people ask
How do I know a subagent is the original? The catalogue marks 33,606 originals and 0 near-duplicates. Prefer an original when you need the clearest source, then inspect the repository file named on its page.
Does a subagent that waits cost anything before it is called? No. Its definition waits until loaded, and the catalogue has 1,782 on-demand entries. Choose this shape for work that appears only when a matching need arrives.
Does an A grade mean I should trust it? No. The scan grade is one signal; 105 entries have been reviewed and 0 are marked high review risk. Read the evidence line before allowing the worker to act.
Do it now
Run npx agentmods add <slug> for a chosen worker, then browse the subagents catalogue.