AI Flow — Multi-Agent Orchestration

Define autonomous AI workflows. Version them like code. Get consistent, repeatable results.

AI Flow adds @flow to your Copilot Chat — a participant that runs multi-role AI workflows defined in .flow.yaml files. Each role gets its own system prompt, tools, and context. Roles run autonomously through full tool-use loops, then hand off to the next role. The result: agents that work longer, stay on-task, and produce consistent output you can trust.

# code-review.flow.yaml — owned by your team, versioned in your repo
name: Code Review
roles:
  - name: Logic & Correctness
    prompt: Review for logical errors, edge cases, and correctness.
    tools: [copilot_readFile, copilot_findTextInFiles]
  - name: Style & Security
    prompt: Review for code style violations and OWASP Top 10 security issues.
  - name: Verdict
    prompt: Synthesize findings into a prioritized, actionable review.

@flow #file:code-review.flow.yaml

Review the changes in src/auth/login.ts

Single-prompt chats hit a ceiling fast. You switch contexts, repeat yourself, lose thread. The model forgets constraints. Output drifts.

AI Flow gives you three things single-prompt chats can't:

Three execution patterns — pick the one that matches your workflow structure.

Roles run in sequence. Each sees the output of every previous role. Best for multi-lens review, staged generation, synthesis workflows.

Security Reviewer  ──▶  Performance Reviewer  ──▶  Lead Engineer (Synthesis)

Stages loop until quality converges. Set a max iteration count per stage. Best for refinement, editing passes, incremental improvement.

stages:
  - name: Draft & Refine
    iterations: 3
    roles:
      - name: Writer
      - name: Critic
  - name: Final Polish
    iterations: 1
    roles:
      - name: Editor

Parallel investigations merged by a synthesizer. Best for incident triage, multi-perspective analysis.

groups:
  - name: App Layer
    roles: [App Investigator]
  - name: Infra Layer
    roles: [Infra Investigator]
  - name: Data Layer
    roles: [Data Investigator]
join:
  name: Incident Commander
  prompt: Synthesize findings into a root-cause report.

Roles don't just chat — they act. Each role gets its own tool allowlist:

tools:
  - copilot_readFile          # Read workspace files
  - copilot_findTextInFiles   # Search the codebase
  - copilot_runInTerminal     # Execute shell commands
  - copilot_listDir           # Browse directory structure
  # or "*" for all available tools

Roles autonomously decide which tools to call, run them, and incorporate results — up to 15 rounds per role. Omit tools for conversation-only roles.

For long-running flows, add delegate: true to a role. It hands off to a background agent via the GitHub Copilot SDK, freeing your chat session while the work completes.

Prompt rendering uses progressive disclosure — lower-priority elements drop first when the token budget is tight, so critical instructions always survive:

This means your flows produce consistent output regardless of how much context you attach.

Production-ready flows ship with the extension. Use them as-is or copy to your workspace and customize.

@flow #file:code-review.flow.yaml

Review the changes in src/auth/login.ts

Three specialized reviewers run in sequence. You get a prioritized review.

No configuration needed. The built-in library is available immediately after install.

• VS Code 1.85+
• GitHub Copilot Chat (pre-installed in VS Code)
• Feima Copilot More LLMs (bundled extension pack — enhanced model selection)

See DEVELOPMENT.md for build, architecture, and release instructions. See AGENTS.md for the AI assistant reference.

MIT