CLAUDE CODE

Parallel work and connected tools

Give each parallel job its own branch or worktree. A connected tool server is real access. Add one only when you will actually use it.

Figure 1: Each parallel job has one owner and its own branch. A connected tool server is real access.

1. Parallel work multiplies the review, not just the speed

A subagent is a focused helper with its own context. A narrow job, such as tests or a review, often goes better there than in a chat that already holds the whole design argument. Your main session stays cleaner.

The bottleneck moves to you. Every extra agent is another diff. Do not run more than you can supervise. Hours of unmerged parallel work produce conflicts and mismatched assumptions. Integrate small pieces often.

2. One owner, and a place the edits cannot collide

Tell each agent which files it owns and which interfaces it must not redesign. Isolation matters more than the count. Give each one its own git branch or worktree, so two writers do not share one folder.

This page does not freeze the flag. Worktree commands change. The current Claude Code docs, and git’s own, are the syntax. The rule that does not change: parallel edits need separate copies.

3. A connected tool is not a longer prompt

Model Context Protocol servers can expose a database, an issue tracker, docs, or a browser. A specialized tool often beats another paragraph of instructions. It also adds context cost and risk. Connect what you will actually use this month, and read the permission you are granting.

Claude Code already works with git, test runners, linters, and package managers. Put those commands in the memory file before you add a new server. Each third-party server is a dependency. It can run code and reach local files, environment variables, and keys. Read it before you install it. Pin a version where you can. Review the update.

Layman rule: A tool you have not read is a person you have not met, with a key to the office.

4. Give it something real to look at

Existing behavior, fixtures, logs, and screenshots beat a description from memory. A browser tool pointed at a page you can also open is checkable. A browser tool pointed at production, with broad permission, is a different decision. Make that decision on purpose.

Instructions hidden in a web page, a ticket, or a tool result can try to steer the agent. That is prompt injection. Be wary of a skill or a server that fetches remote content, asks for broad permission, or runs a command you cannot read.

5. What to try next

If you try two agents, give them two branches and two file lists, and merge the first one before you start a third. List the MCP servers you already installed and delete the one you have not used.

Access, tests, and a small set of real tasks are the next habit: permissions, checks, and a small eval set.