Claude's helpers: subagents and Remote Control

A crowded conversation

Kai asks Claude Code: "find every place Hyper Shards uses laserDamage and check the tests still cover it". Claude opens 40 files, and every one of them lands in the conversation.

Everything the model can see at once is its context window, a bit like its desk. The desk has a size limit. Cover it with 40 files Kai will never look at again, and there is less room, and less focus, for the real job.

Big jobs need a way to keep the desk clear.

Hand it to a helper

A subagent is a helper agent that the main agent sends off to do one focused job:

• The main agent writes a short job description.
• The subagent starts with its own fresh context window. It doesn't see your conversation, only that job description.
• It does the heavy work: opens the 40 files, runs the searches.
• It sends back only a summary, like "laserDamage is used in 3 files, and 2 tests cover it".

The 40 files stay on the helper's desk. Kai's conversation gets the answer.

Why helpers help

• A clear deskSearch results, logs and long files stay in the helper's context. The main conversation stays focused.
• Several at onceBackground subagents run at the same time while you keep working. One checks the Hyper Shards tests while another lists every pest in Garden Defense.
• SpecialistsClaude Code has built-in helpers. Explore searches code and can't edit anything. Plan researches during plan mode. A general-purpose helper can explore and act.
• Not freeEach helper sends its own requests to the model, and they count toward the same usage limits. Use helpers for big jobs, not one-line questions.

Make your own helper

A studio can define its own helpers. Each one is a Markdown file (plain text with simple formatting) in the repo's .claude/agents/ folder, so the whole team gets it through Git.

The top of the file, between two --- lines, is the frontmatter: the helper's settings.

• name and description are required. Claude reads the description to decide when to call this helper.
• tools lists exactly which tools it may use. Leave it out and the helper gets every tool subagents can have.

Below the frontmatter come the helper's own instructions.

CLAUDE.md: notes for every session

Every Claude Code session starts with an empty context window, so Claude doesn't remember yesterday. Explaining "numbers live in balance.json" every morning gets old fast.

So the repo has a file called CLAUDE.md. Claude Code reads it at the start of every session: build commands, rules, where things live. For example: "Run npm test before every commit." It is committed to Git, so the whole team shares it, and our own helpers read it too.

Keep it short (Anthropic suggests under 200 lines) and concrete. /init writes a first draft by studying the code.

Bolt says: Claude makes the same mistake twice? That is the moment to add a line to CLAUDE.md.

Remote Control

Kai starts a long job in Claude Code, then has to leave for football. Typing /remote-control connects the session. On the phone, Kai opens the Claude app's Code tab (or claude.ai/code in a browser) and picks the session.

Claude keeps running on Kai's laptop, with its commands and files. The phone is a window: Kai reads progress, answers questions and approves actions from the sideline.

So the laptop must stay on with Claude Code running. If it sleeps, the session reconnects when it wakes. The laptop only makes outgoing connections, and never opens a door for others to connect in.

Who says yes?

Helpers working in parallel and a session steered from a football pitch raise a real question. When Claude wants to delete a folder or push to GitHub, and nobody is at the keyboard, who decides?

Claude Code's answer is permissions: settings that say what Claude may do alone and what needs a human yes.

Bolt says: Next lesson: permission modes, and how a studio picks the right one for each job.