Skip to content

agent-designer.md

Latest commit

 

History

History
228 lines (178 loc) · 8.71 KB

agent-designer.md

File metadata and controls

228 lines (178 loc) · 8.71 KB
Raw
description Agent Designer is a built-in agent that generates valid AgentSpec YAMLs from natural language. Available immediately after installing the extension.

Agent Designer

Status: Coming soon

@agent-designer is a built-in agent bundled with the extension. It is available immediately after installation — no marketplace copy or manual setup required.

Agent Designer

The Agent Designer reads your workspace context — existing agents, installed skills, and your project's technology stack — and produces a spec that fits naturally into your team architecture. It:

  • Infers the correct role, domain, expertise, and intents from your description
  • Suggests up to 3 relevant workspace skills and adds them to skills with clear when conditions
  • Avoids collisions with existing agent IDs
  • Only references skills that are actually installed in the workspace
  • Emits exactly one fenced YAML block — no explanatory text, no placeholders

It also supports importing existing agent definitions from other formats (Claude Code .md, Copilot .agent.md, plain markdown) and converting them to a valid AgentSpec YAML.

Using the Agent Designer

Step 1 — Invoke the agent

The invocation syntax depends on your AI tool:

  • GitHub Copilot Chat — use the @agent-designer chat participant: 
    @agent-designer design an agent that reviews TypeScript pull requests for security issues and missing tests
  • Claude Code — use the /agent-designer slash command: 
    /agent-designer design an agent that reviews TypeScript pull requests for security issues and missing tests

Step 2 — Describe what you need

The more context you give, the more accurate the result. You can include:

Hint Example
Role "an orchestrator that coordinates the frontend pipeline"
Domain "a backend agent focused on Node.js REST services"
Main capability "reviews PRs", "generates API docs", "runs database migrations"
Constraints "read-only, no file creation"
Skills to use "use the search-codebase skill"

Tip: You don't need to mention every field. The agent infers what it can from the workspace and leaves optional fields empty when unsure.

Step 3 — Use the generated YAML

The agent outputs a single YAML code block. You can:

  1. Copy → Create Agent wizard — paste the YAML into the Raw YAML tab inside the agent wizard
  2. Save directly — place the file in .agent-teams/agents/your-agent.yml and the dashboard will detect it on the next refresh
  3. Iterate — ask the agent to adjust specific fields: refine the constraints to also escalate when modifying a public API

Importing an existing agent definition

Provide the content of an existing agent file from any supported format and the Agent Designer will convert it:

# Copilot
@agent-designer convert this Claude Code agent to AgentSpec YAML:
<paste the .md file content here>

# Claude Code
/agent-designer convert this Claude Code agent to AgentSpec YAML:
<paste the .md file content here>

The agent extracts every mappable field, normalises values to AgentSpec conventions (role names, intent snake_case, permissions), and asks a single grouped clarification message for anything it cannot resolve before emitting the YAML.

Example

Input:

# Copilot
@agent-designer create a worker agent that reviews pull requests in a TypeScript + Node.js backend project

# Claude Code
/agent-designer create a worker agent that reviews pull requests in a TypeScript + Node.js backend project

Output:

id: pr-reviewer
name: PR Reviewer
version: 1.0.0
role: worker
domain: backend
description: >
  Reviews pull requests for correctness, security, and test coverage in
  TypeScript Node.js services. Leaves structured comments and approves or
  requests changes.
expertise:
  - TypeScript
  - Node.js
  - REST API design
  - security code review
  - test coverage analysis
intents:
  - review_pull_request
  - check_test_coverage
  - flag_security_issues
scope:
  topics:
    - pull request review
    - code quality
    - security analysis
  path_globs:
    - pattern: src/**/*.ts
      priority: high
    - pattern: '**/*.test.ts'
      priority: medium
  excludes:
    - dist/**
    - node_modules/**
workflow:
  - Read the PR diff and identify changed files within scope.
  - Check for missing or inadequate tests for changed logic.
  - Scan for common security issues (injection, auth bypass, secrets in code).
  - Review API contracts for breaking changes.
  - Produce a structured review using the code-review output template.
tools:
  - name: github
    when: fetching PR diff and posting review comments
permissions:
  can_create_files: false
  can_edit_files: false
  can_delete_files: false
  can_run_commands: false
  can_delegate: false
  can_modify_public_api: false
  can_touch_global_config: false
constraints:
  always:
    - Flag any change that removes or weakens authentication checks.
    - Require tests for every new exported function.
  never:
    - Approve a PR that introduces hardcoded secrets or credentials.
    - Review files outside the defined scope.
  escalate:
    - When a PR modifies public API contracts without a version bump.
output:
  template: code-review
  mode: detailed
  max_items: 10
targets:
  - github_copilot
  - claude_code

The agent-spec-authoring skill

The agent-spec-authoring skill is bundled with the extension — no manual installation or workspace setup is required. It is injected automatically into every @agent-designer session. The skill provides the complete AgentSpec schema — all valid fields, types, validation rules, and role semantics — so the agent always produces specs that pass validation.

Local override: if a skill with id agent-spec-authoring is present in .agent-teams/skills/agent-spec-authoring/, it takes precedence over the bundled version automatically.

Required fields

Field Type Rules
id string Slug: ^[a-z0-9-]+$. Must be unique in the workspace.
name string Human-readable, 3–80 characters.
role string worker | orchestrator | router | aggregator
description string 1–3 sentences, 10–600 characters.

Role semantics

Role When to use
worker Executes a focused task within a defined scope
orchestrator Decomposes a goal and delegates sub-tasks to workers within a domain
router Receives requests and routes them to the best-matched agent via dispatch tools
aggregator Merges results from parallel orchestrators and surfaces cross-domain conflicts

Common optional fields

Field Description
domain Primary domain: frontend, backend, testing, tooling, devops
expertise List of knowledge areas used for routing
intents Snake_case action IDs this agent handles (e.g. review_pull_request)
scope File globs, topics, and excludes that define the agent's operational boundary
workflow Ordered imperative steps the agent follows
skills Skill IDs from the workspace registry
tools Environment capabilities (github, terminal, etc.) with optional when
constraints always / never / escalate behavioral rules
targets Platforms to sync to: copilot, claude (default: both)

For the full schema reference including sub-object definitions, see the SKILL.md inside marketplace/skills/agent-spec-authoring/ or the bundled copy in the extension.

Constraints

The Agent Designer enforces these rules in every generation:

Always:

  • Emit exactly one fenced YAML block — no text before or after
  • Use only fields defined in the AgentSpec schema
  • id must match ^[a-z0-9-]+$ and must not collide with existing agent IDs
  • Every intent must match ^[a-z0-9_]+$ (snake_case)
  • Include all 4 required fields: id, name, role, description
  • Reference only skills available in the current workspace

Never:

  • Invent field names not in the schema
  • Emit JSON instead of YAML
  • Include markdown prose outside the fenced block
  • Reuse an ID that already exists in the workspace