microsoft/amplifier-openclaw

Amplifier — Multi-Agent Delegation

Amplifier is a multi-agent AI framework. Delegate tasks that benefit from specialist agents, structured workflows, or parallel investigation.

When to Delegate

High confidence → delegate immediately:

• "Research X and compare approaches"
• "Build a Python tool that does X"
• "Review this code for security and design"
• User says "amplifier", "deep dive", "thorough", "comprehensive"
• Task has clear subtasks benefiting from parallel agents

Medium confidence → offer the choice:

• "I can do a quick analysis, or delegate to Amplifier for a thorough multi-agent review."

Low confidence → handle yourself:

• Simple Q&A, quick code edits, casual conversation, anything needing immediate response

Usage

Basic Delegation

exec command:"amplifier-openclaw run 'Research the top 3 Python web frameworks' --bundle foundation" background:true timeout:600

With Model Selection

Pass --model to override the default model:

exec command:"amplifier-openclaw run --model your-preferred-model 'Deep code review' --bundle foundation" background:true timeout:600

Tip: Pass --model matching the model from your OpenClaw Runtime line so Amplifier uses the same one.

Bundles

amplifier-openclaw bundles list

| Bundle | Best For | |--------|----------| | foundation | General: research, analysis, planning (default) | | superpowers | Multi-agent brainstorm, deep investigation | | coder | Code generation, refactoring, debugging |

Session Persistence

# Start a named session
exec command:"amplifier-openclaw run --session-name my-project 'Start building the auth module' --bundle foundation" background:true

# Resume later
exec command:"amplifier-openclaw run --resume --session-name my-project 'Now add unit tests'" background:true

Modes

Amplifier supports slash-command modes in prompts. Modes do not carry over between runs — include the mode at the start of each prompt:

# Brainstorm mode (uses all agents)
exec command:"amplifier-openclaw run --bundle superpowers '/brainstorm How should we architect the new API?'" background:true

# Research mode
exec command:"amplifier-openclaw run --bundle foundation '/research Latest advances in RAG'" background:true

JSON Output

{
  "response": "The analysis found...",
  "usage": {
    "input_tokens": 28566,
    "output_tokens": 1800,
    "estimated_cost": 0.12,
    "tool_invocations": 3
  },
  "status": "completed"
}

Cost Tracking

exec command:"amplifier-openclaw cost --period week"

Report costs only when asked or when notable (>$1).

Interpreting Results

• response: Present to the user (the main output)
• error: Report in plain language, don't dump raw JSON
• usage.estimated_cost: May be 0.0 — don't alarm about zero
• status: "completed", "cancelled", or error state

During Active Delegation

• "stop"/"cancel" → kill the background process
• Unrelated questions → answer yourself, don't interrupt Amplifier
• Follow-up → tell user you'll pass it along when current task finishes

Install

If not already installed:

uv tool install "amplifier-app-openclaw @ git+https://github.com/microsoft/[email protected]"