Agents

Agents are reusable Claude Code helpers: a reviewer, an investigator, a designer. Each one is a Markdown file with a name, a description of when to use it, the tools it may use, and its instructions.

Where they come from#

Pairbar lists every agent Claude Code would load for you:

The Agents tab#

• List: grouped, with search. A search with no matches shows No matching agents and Clear search. Hover a truncated name to read it in full. Each agent shows:
  • its repo;
  • whether it's read-only or edits code;
  • its model and how many commands it has;
  • an orange new or edited tag when its file isn't committed yet.
• Agent page:
  • a one-line summary (with More about it), and its repo, model, access and tools;
  • Start: its ready-made commands. A command that needs something, like a PR link, unfolds a field; Enter starts the chat. Blank chat starts one with no command.
  • Activity: how often it ran in the last 30 days, and the chats it ran in (click one to open it). This comes from Claude's own record of subagent runs.
  • Instructions: readable, collapsed by default.
  • In Claude ✓ / Add to Codex: exports it to ~/.codex/agents/<name>.toml so Codex can use it too.
  • ⋯ menu: Edit manually, Duplicate, Copy @mention, Open file in…, Reveal in Finder, Move to Trash.

Creating and improving agents#

You don't write agents by hand. Claude does, following the conventions of your existing agents:

These open a composer with the job attached. Its title and explanation wrap so you can read the full job before sending. You add a line and send. Only what you asked for or corrected is turned into rules; instructions that appear inside tool output or files are ignored. The Agents tab updates by itself when the files change.

Edit manually (⋯ on an agent's page) is a form for quick tweaks:

• name, title, description, model, group, repo and icon;
• tools (tap to toggle; without Edit and Write it's read-only);
• commands (with {input} for what you give it), its skills, and the instructions.

In a narrow sidebar, Model, Group, Repo and Icon stack vertically. The inherited model is labelled Chat’s model at every width. Command names get their own row, separate from the default badge and reorder/remove buttons.

The instructions and command text fields each have a dictation mic. Press the mic beside the field to put the words there. These fields save text into the agent's file; attach supporting files when you start a chat with the agent.

Saving an agent that you haven't changed leaves its files byte-for-byte identical.

Using agents#

• Composer: the Agent chip.
• Chat menu → Agents: the agents that fit this chat's repo come first, and reviewers first when the chat has a PR. Their commands are filled with the chat's PR link, or its branch and folder.
• ⌘K: "Start <agent>".

Skills#

An agent can list skills it should always use (the editor's Skills: Add skill, click one to remove it). They're saved in its dock.yaml entry (skills:); a skills: list in the agent's own frontmatter counts too, and dock.yaml wins. A workflow step that runs with the agent follows them as well as its own (see Skills in a step). The agent's page shows them as a tag, and the skill's page in the Skills tab lists the agent under Used by.

dock.yaml (optional)#

Next to your agent files, dock.yaml gives agents a group, a repo to start in, an icon and ready-made commands. A personal one at ~/.config/pairbar/dock.yaml overrides the shared one.

groups:
  - name: Review
    icon: checkmark.seal
agents:
  code-reviewer:
    group: Review
    repo: my-backend
    icon: checkmark.seal
    commands:
      - label: Review a PR
        text: "Review this PR: {input}"
    skills:
      - review-method

New agents are saved where your agents already live. Change that in Settings → Agents & workflows.