Delegate complex tasks to Amplifier's multi-agent framework. Use when: (1) research/comparison needing multiple perspectives, (2) multi-file code projects, (3) architecture/design reviews, (4) user asks for deep/thorough work. NOT for: simple Q&A, quick edits, casual chat, anything needing <5s response. CLI: amplifier-openclaw.
Amplifier is a multi-agent AI framework. Delegate tasks that benefit from specialist agents, structured workflows, or parallel investigation.
High confidence → delegate immediately:
Medium confidence → offer the choice:
Low confidence → handle yourself:
exec command:"amplifier-openclaw run 'Research the top 3 Python web frameworks' --bundle foundation" background:true timeout:600
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.
amplifier-openclaw bundles list
| Bundle | Best For |
|---|---|
foundation | General: research, analysis, planning (default) |
superpowers | Multi-agent brainstorm, deep investigation |
coder | Code generation, refactoring, debugging |
# 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
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
{
"response": "The analysis found...",
"usage": {
"input_tokens": 28566,
"output_tokens": 1800,
"estimated_cost": 0.12,
"tool_invocations": 3
},
"status": "completed"
}
exec command:"amplifier-openclaw cost --period week"
Report costs only when asked or when notable (>$1).
response: Present to the user (the main output)error: Report in plain language, don't dump raw JSONusage.estimated_cost: May be 0.0 — don't alarm about zerostatus: "completed", "cancelled", or error stateIf not already installed:
uv tool install "amplifier-app-openclaw @ git+https://github.com/microsoft/[email protected]"