夜雨聆风We can direct an AI coding agent to read our codebase, make edits, run commands, fix its own mistakes, and accomplish work. These agents operate in a loop. We give direction, the agent acts, we course-correct, and repeat until we reach our goal.
Custom agents let the AI adopt specialized personas with specific tool access. For example, we can create a code reviewer agent that can only read code, not modify it.
• Custom agents allow us to configure AI to take on specific roles. • Custom agents allow us to define a persona with a clear purpose, behavior, and scope.
We can define a persona like security reviewer, planner, or solution architect, instead of a generic assistant. Without custom agents, we’re only relying on a single, general-purpose assistant for everything.
With custom agents, we can switch between specialized roles depending on the task.
• more focused and relevant outputs • better alignment with specific tasks • clearer structure in responses
Instead of constantly re-explaining context, we select the right agent and let it handle the task.
Type /create-agent, followed by a description of the agent we want to create, and press Enter.
/create-agent A code reviewer that analyzes code for quality, security, and best practices without modifying filesWe can also ask the generic agent to suggest what custom agent should we create:
Based on my current project. What custom agent can I create for my video demo? Can you tell me a prompt to tell Copilot to create one? And tell me why it makes a good custom agent?Note that this approach helps us think beyond just writing code and start designing workflows, and we can see that Copilot sugguests an agent designed specifically for an arcade-themed calculator project.
Custom agents give us structured, role-based control over AI. Here the custom agent in this example is built with:
• awareness of the project’s architecture • knowledge of UI themes and styling • consistency with coding patterns
When we ask it to build something, like a new calculator app, it applies all of its predefined rules and context automatically. For example, we can ask this custom agent to create a tip calculator.
What’s interesting is that the new app inherits characteristics from the original project, including:
• visual style and theme • sound effects • structural patterns
The custom agent ensures consistency without needing to restate requirements.
Each agent can have its own to transform the AI from a general assistant into a role-specific collaborator.
• instructions and behavior • access to tools • understanding of our project
We should answer the clarifying questions about the agent and review the generated agent file, then wen can select the Reviewer agent from the agents dropdown in the Chat view, then ask "review the authentication module". The agent should analyze the code without making changes.
Q: Should this reviewer focus on all files or specific types?A: All files - Full-stack code reviewQ: How detailed should the review be?A: Comprehensive - Quality, security, performance, best practicesQ: How should results be delivered?A: In chat as bullet list - Issues grouped by severityQ: Should it be able to run terminal commands (e.g., linters)?A: Read-only (no execute) - Search + read only, safe for any codebaseNow I have all the info:- Scope: All files- Depth: Comprehensive (quality, security, performance, best practices)- Output: In chat as bullet list grouped by severity- Tools: Read-only (search, read)- Should be user-invocable: true (so user can select it from the agent picker)Let me also check - the existing react-standards.instructions.md could be referenced. Let me create the agent file.Let me also check the existing files structure.- Read .githubGood, .github/agents/ doesn't exist yet. Let me create it along with the agent file.- Created agentsNow let me create the agent file.- Created code-reviewer.agent.mdLet me validate the file - check frontmatter syntax, description, tools, etc.- Read code-reviewer.agent.mdThe agent file looks good. Let me present the results to the user.One thing I want to note - the model Claude Sonnet 4.5 (copilot) might not be the exact model name. But since the user's environment info says they are on macOS and this is VS Code, the model names would be the ones available in their VS Code Copilot setup. This should be fine.Let me now present the summary.Agent file validated — frontmatter, tools, and constraints are correct. Here's the summary:| File | Description ||------|-------------|| code-reviewer.agent.md | Read-only code review agent that analyzes quality, security, performance, and best practices without modifying any files. |**What this agent does:**- Analyzes code across 5 dimensions: correctness, security, performance, best practices, maintainability- Cross-references project conventions (AGENTS.md, instructions)- Returns findings grouped by severity (🔴 Critical / 🟡 Warning / 🔵 Suggestion) with file paths, line numbers, explanations, and fix suggestions-**Read-only** — only `read` and `search` tools, no file modifications or terminal access**Example invocations:**-`@Code Reviewer Review the current branch changes` — reviews the git diff-`@Code Reviewer Review src/components/LoginForm.tsx` — reviews a specific file-`@Code Reviewer Review the selected code` — reviews your current selection**Related customizations you might want next:**-`/create-hook pre-commit-lint ...` — A `PreToolUse` hook that runs the linter before any edit tool is used-`/create-instructions security-guidelines ...` — File-specific instructions for security-sensitive code (auth, API keys, SQL)-`/create-prompt analyze-api ...` — A prompt that invokes the Code Reviewer agent specifically on API endpoints- Add the Code Reviewer as a subagent reference in AGENTS.md so the default agent can delegate review tasks automaticallyMade changes.We can add handoffs to our agent to create guided workflows. For example, hand off from a planning agent to an implementation agent.
The custom agent can define reusable roles that carry context, behavior, and intent across tasks, we are not no longer need to start from scratch every time, we are working with systems that understand our project and apply that understanding consitently.
Custom agents represent a shift from:
single assistant → team of specialists
We can create multiple agents, each optimized for a specific role, instead of one AI do everything.
