Contents

Features

Agent Configurator

•7 min read
Agent Configurator

Agent Configurator

If you have ever wanted to change an agent's model or lock its tools to read-only and ended up staring at YAML, this is for you. Open any agent file in Ritemark and the right side of the screen switches to the Agent Configurator — a visual editor built on the real Claude Code agent format. What you see is exactly what the AI runtime reads.

Agent Library on the left and the Agent Configurator on the right, with an agent file open in the middle

There is no separate save button. Every change writes straight into the agent's markdown file, so every edit shows up in git diff and can be reviewed or reverted like any other change to your repo.

How to Open It

Click the robot icon in the left activity bar to open the Agent Library, then click any agent file. The editor opens the agent's instructions in the center, and the right panel becomes the Agent Configurator.

The top toolbar carries three panel toggles. Click them to switch between views:

Icon Panel What it does
List icon Table of Contents Navigate long instruction documents
Info icon Properties Raw view of every settings field
Robot icon Agent Configurator The visual editor described below

Instructions files (CLAUDE.md, AGENTS.md) live in their own section of the library. They are project-wide rules loaded into every AI session — house rules, not team members — so they intentionally do not get a Configurator panel.

The Fields, Top to Bottom

Description (required)

This is the most important field. It tells the AI when to hand work to this agent. Write it like a job posting. A good Description for a PR-review agent might read: "PR review and merge agent. Invoke when user mentions: review PR, merge PR, check PR."

If the description is empty, the agent never gets invoked automatically — which is why the Configurator shows a warning until you fill it in. You can still launch the agent manually from the library, but automatic delegation will not happen.

Model

Which AI model the agent runs on:

  • Inherit (default) — uses whatever model the session is using. This is the right default for almost everything.
  • Sonnet — fast and capable. Good for routine work.
  • Opus — most capable. Good for complex reasoning.
  • Haiku — fastest and lightest. Good for simple, high-volume tasks.
  • Custom model ID — pin an exact model version when you need one.

Tools (allow-list)

The Tools section controls what the agent is allowed to touch. It works as an allow-list with one important rule:

  • Nothing checked = inherits everything. The agent can use every tool your session can use. The Configurator labels this state explicitly so you do not mistake an empty list for "no tools."
  • Some tools checked = only those tools. An unchecked tool simply does not exist for that agent. It cannot read the web, cannot run shell commands, cannot write files unless you have granted it.
Tool What the agent can do
Read Read file contents
Write Create and overwrite files
Edit Make targeted edits to files
Bash Run shell commands
Glob Find files by pattern
Grep Search file contents
WebFetch Fetch and read web pages
WebSearch Search the web
NotebookEdit Edit Jupyter notebooks
Skill Invoke skills

A practical example. A "Product Marketer" agent with only Read, Write, Edit, Glob, and Grep checked can work with files in your repo but can never browse the web, never run a shell command, and never reach outside the workspace. That is least privilege by design.

Tip: For a research agent that should never modify anything, check Read, Glob, Grep, WebFetch, and WebSearch. It can investigate, never edit.

Skills

Preload Ritemark skills into the agent's context so every run starts knowing those workflows. Type to search your project and user skills, click to add, click a tag to remove. Skills are a separate concept from agents — they are reusable workflow snippets the agent can call.

Advanced

Collapsed by default. Open it for:

  • Effort — how much reasoning effort the agent applies, from low to max. Higher effort costs more tokens but produces more careful work.
  • Memory — give the agent persistent memory across sessions, scoped to user, project, or local.
  • Color — the agent's display color in lists and UI chips.

Linked Flow (Ritemark extension)

Attach a Ritemark Flow to this agent. This field is a Ritemark-specific addition on top of the standard agent format, and the panel marks it as such. Use it when you want an agent to trigger a defined visual workflow.

Saving and Version Control

Every field in the Configurator writes immediately to the agent's .md file. There is no save button because there is nothing to save — the file on disk and the visual state are always the same thing.

That means:

  • All changes appear in git status and git diff like any other file change.
  • You can revert with git checkout -- .claude/agents/my-agent.md.
  • You can review configuration changes in a pull request like any other code change.
  • Two people editing the same agent file go through the normal git merge process.

Writing the Agent's Instructions

The document body below the settings is the agent's system prompt — the instructions it follows every time it runs. Edit it in the normal Ritemark editor. Good instructions usually cover four things:

  1. Who the agent is. "You are a UX/UI design specialist for..."
  2. What the agent must always do. Core principles, output format, tone.
  3. What the agent must never do. Guardrails.
  4. Examples. The fastest way to teach an agent your standards is to show a few input-output pairs.

Three Quick Recipes

Research agent that cannot touch your files. Tools: Read, Glob, Grep, WebFetch, WebSearch. It can investigate but never modify.

Writing agent locked to your repo. Tools: Read, Write, Edit, Glob, Grep. No Bash, no web. It edits documents but cannot run anything or leak attention to the internet.

Heavyweight reviewer. Model: Opus. Effort: high. Tools: Read, Bash, Glob, Grep. Maximum scrutiny, no write access.

Using an Agent

Once an agent is configured, you can use it three ways:

  • Automatic delegation. When your request matches the agent's Description, the AI hands the task to that agent without you doing anything.
  • Launch Chat. Right-click any agent in the library and pick Launch Chat to start a conversation with that agent pinned.
  • @ mention. Type @agent-name in the AI sidebar composer to route work to it explicitly.

Where Things Live

Thing Location Shared via git?
Project agents .claude/agents/*.md Yes
Personal agents ~/.claude/agents/*.md No
Project skills .claude/skills/<name>/SKILL.md Yes
Codex skills .agents/skills/<name>/SKILL.md Yes
Slash commands .claude/commands/*.md Yes
Project instructions CLAUDE.md, AGENTS.md (repo root) Yes
Flows .ritemark/flows/*.flow.json Yes

Quick Reference

Action How
Open the Configurator Click an agent in the Agent Library, then the robot toggle on the right toolbar
Switch to the table of contents List icon on the right toolbar
See raw settings Info icon on the right toolbar
Make an agent read-only Tools: check only Read, Glob, Grep
Grant an agent all tools Leave every tool box unchecked
Make an agent never delegate Leave the Description empty (warning shown)
Revert a configuration change git checkout -- .claude/agents/<name>.md
Back to support