Skip to main content

Agents Overview

Copilot Collections provides 13 user-facing specialized agents plus internal worker agents that together form an AI product engineering team covering the full delivery lifecycle — from product ideation through development, infrastructure, and quality assurance. Agents are stored in .github/agents/ as .agent.md files. VS Code loads the user-facing agents automatically when the corresponding mode is selected.

How Agents Work

Each agent has:

  • A defined role — What the agent specializes in and what it should/shouldn't do.
  • Tool access — Which MCP integrations and VS Code tools it can use.
  • Skill bindings — Which skills it loads for domain-specific knowledge.
  • Handoffs — Buttons to seamlessly transition between workflow phases.

Agent Handoff Diagram

┌──────────────────────┐
│ Business Analyst │
│ /tsh-analyze-materials│
└──────┬───────────────┘
│ Start Implementation

┌─────────────────────────┐
│ Engineering Manager │ ← Orchestrates the full cycle
│ /tsh-implement │
└──────┬──────────────────┘
│ Delegates to specialized agents
├─────────────────┬─────────────────┬─────────────────┬─────────────────┬─────────────────┬─────────────────┬─────────────────┬─────────────────┬─────────────────┬─────────────────┬─────────────────┐
▼ ▼ ▼ ▼ ▼ ▼ ▼ ▼ ▼ ▼ ▼
┌──────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ Context │ │ Architect │ │ Architect │ │ UI │ │ Software │ │ Plan │ │ DevOps │ │ E2E │ │ Prompt │ │ UI Reviewer │ │ Technical │
│ Engineer │ │ (plan) │ │ Reviewer │ │ Engineer │ │ Engineer │ │ Implementor │ │ Engineer │ │ Engineer │ │ Engineer │ │ /tsh-review-│ │ Writer │
│ (research) │ │ │ │ (plan review)│ │ (UI) │ │ (non-UI) │ │ (strict │ │ (infra) │ │ (tests) │ │ (prompts) │ │ ui │ │ (docs) │
└──────────────┘ └──────────────┘ └──────────────┘ └──────────────┘ └──────────────┘ └──────────────┘ └──────────────┘ └──────────────┘ └──────────────┘ └──────────────┘ └──────────────┘


┌──────────────┐
│ Code Reviewer │
│ /tsh-review │
└──────────────┘

Both Architect handoffs — Start Implementation and Start Infrastructure Implementation — pass through the Engineering Manager. The manager owns the Human approval gate for the exact current plan revision before the first file-changing delegation; an automated Reviewer approval does not authorize implementation.

All seven execution owners apply the same inline, fail-closed precondition before editing: they read the persisted Human Approval record from disk, name the exact failed field, condition, or file when validation fails, and use vscode/askQuestions for guided recovery on both delegated and direct entry paths. A delegated owner may offer hand-back to tsh-engineering-manager as one choice, but does not dead-end there. The Engineering Manager and Architect present and record Human Approval, tsh-plan-reviewer provides non-authorizing Reviewer approval, and execution owners validate before edits.

Agent Summary

📋 Product Ideation Agents

AgentFileRoleKey Tools
Business Analysttsh-business-analyst.agent.mdOrchestrates workshop analysis into Jira-ready epics and storiesAtlassian, Figma, PDF Reader, Sequential Thinking

🛠 Development Agents

AgentFileRoleKey Tools
Context Engineertsh-context-engineer.agent.mdGathers requirements, builds context, identifies gapsAtlassian, Figma, PDF Reader, Sequential Thinking
Architecttsh-architect.agent.mdDesigns solutions, creates implementation plansAtlassian, Context7, Figma, PDF Reader, Sequential Thinking
Engineering Managertsh-engineering-manager.agent.mdOrchestrates implementation by delegating to specialized agentsAtlassian, Sequential Thinking
UI Engineertsh-ui-engineer.agent.mdImplements UI and frontend solutionsContext7, Figma, Playwright, Sequential Thinking
Software Engineertsh-software-engineer.agent.mdImplements non-UI code against the planContext7, Sequential Thinking
Prompt Engineertsh-prompt-engineer.agent.mdDesigns, optimizes, and secures LLM application promptsContext7, Sequential Thinking

🏗 Infrastructure & DevOps Agents

AgentFileRoleKey Tools
DevOps Engineertsh-devops-engineer.agent.mdInfrastructure automation, CI/CD, cloud governance, cost optimizationContext7, Sequential Thinking, AWS API, AWS Docs, GCP Gcloud, GCP Observability, GCP Storage

✅ Quality Agents

AgentFileRoleKey Tools
Code Reviewertsh-code-reviewer.agent.mdReviews code quality, security, correctnessAtlassian, Context7, Figma, Sequential Thinking
UI Reviewertsh-ui-reviewer.agent.mdVerifies UI matches Figma designFigma, Playwright, Context7
E2E Engineertsh-e2e-engineer.agent.mdCreates and maintains Playwright E2E testsPlaywright, Context7, Figma, Sequential Thinking

⚙️ Copilot Customization Agents

AgentFileRoleKey Tools
Copilot Engineertsh-copilot-engineer.agent.mdDesigns, creates, reviews Copilot customization artifactsContext7, Sequential Thinking
Copilot Orchestratortsh-copilot-orchestrator.agent.mdCoordinates complex multi-step Copilot customization tasksSequential Thinking

🔧 Internal Worker Agents

These agents are not invoked directly by users. The Business Analyst internally delegates transcript, analysis, extraction, quality-review, and formatting phases to BA workers, the Copilot Orchestrator delegates to the customization workers below, and the Engineering Manager delegates implementation-specialist and documentation-only work to the workers below.

AgentFileRole
Copilot Researchertsh-copilot-researcher.agent.mdAnalyzes codebases and documentation, extracts patterns
Copilot Artifact Creatortsh-copilot-artifact-creator.agent.mdCreates and modifies Copilot customization artifacts
Copilot Artifact Reviewertsh-copilot-artifact-reviewer.agent.mdValidates quality and consistency of artifacts
Architect Reviewertsh-plan-reviewer.agent.mdStress-tests implementation plans before implementation starts
Plan Implementortsh-plan-implementor.agent.mdInternal strict single-task implementor for one plan step at a time
Technical Writertsh-technical-writer.agent.mdAuthors and updates README, CHANGELOG, /docs, and published documentation-site pages