AgentPipe

Orchestrate conversations between multiple AI agents. Useful for:

•Getting multiple perspectives on a problem
•Code review from different AI models
•Collaborative brainstorming
•Comparative analysis

Quick Start

Two-Agent Conversation

bash

agentpipe run -a claude:Alice -a gemini:Bob -p "Review this approach to caching"

With TUI and Metrics

bash

agentpipe run -a claude:Reviewer -a codex:Implementer --tui --metrics -p "Design a rate limiter"

From Config File

bash

agentpipe run -c /path/to/config.yaml --tui

Agent Specification

Three formats for -a/--agents:

Format	Example	Description
`type`	`-a claude`	Just the agent type
`type:name`	`-a claude:Alice`	Type with display name
`type:model:name`	`-a claude:claude-sonnet-4-5:Alice`	Type, model, and name

Available Agents in boxctl

These are installed in the container:

•claude - Anthropic Claude
•codex - OpenAI Codex
•gemini - Google Gemini
•qwen - Alibaba Qwen

Configuration File

For complex setups, use a YAML config:

yaml

version: "1.0"

agents:
  - id: reviewer
    type: claude
    name: "Code Reviewer"
    prompt: "You are a senior engineer focused on code quality and bugs."

  - id: architect
    type: gemini
    name: "Architect"
    prompt: "You focus on system design and scalability."

orchestrator:
  mode: round-robin      # round-robin | reactive | free-form
  max_turns: 6
  initial_prompt: "Review this pull request for issues"

logging:
  enabled: true
  show_metrics: true

Conversation Modes

Mode	Behavior
`round-robin`	Fixed rotation: A → B → A → B
`reactive`	Response based on previous speaker's content
`free-form`	Agents decide when to participate

Key Flags

bash

agentpipe run [flags]

  -a, --agents        Agent specs (can repeat)
  -p, --prompt        Initial prompt
  -c, --config        YAML config file
  --mode              Conversation mode
  --max-turns         Max rounds
  --timeout           Per-agent timeout
  --tui               Terminal UI mode
  --no-tui            Disable TUI (for scripting)
  --metrics           Show response metrics
  --json              JSON output

Common Patterns

Code Review Panel

bash

agentpipe run \
  -a claude:Bugs -a gemini:Architecture -a codex:Performance \
  -p "Review this code for issues: $(cat src/api.py)" \
  --mode round-robin --max-turns 3

Architecture Discussion

bash

agentpipe run \
  -a claude:Pragmatist -a gemini:Idealist \
  -p "Should we use microservices or monolith for this startup?" \
  --tui --metrics

Scripted (No TUI)

bash

# For automation - capture output
agentpipe run \
  -a claude -a gemini \
  -p "Summarize the pros/cons of Rust vs Go" \
  --no-tui --max-turns 4 > discussion.txt

Utility Commands

bash

# Check what agents are available
agentpipe doctor

# List agents with details
agentpipe agents list
agentpipe agents list --json

Tips

•
Name your agents - Makes conversations easier to follow: -a claude:Skeptic -a gemini:Optimist
•
Use prompts to set roles - In config files, give each agent a specific perspective
•
Limit turns - Start with --max-turns 4-6 to keep conversations focused
•
TUI for interactive - Use --tui when watching live, --no-tui for scripts
•
Combine with boxctl-analyst - Use agentpipe for brainstorming, analyst MCP for deep dives

When to Use AgentPipe vs Analyst MCP

Use Case	Tool
Multi-perspective discussion	AgentPipe
Deep code analysis	boxctl-analyst
Brainstorming approaches	AgentPipe
Code review with report	boxctl-analyst
Comparing viewpoints	AgentPipe
Plan verification	boxctl-analyst

Troubleshooting

"Agent not found"

Run agentpipe doctor to see which agents are installed.

Conversation stalls

Add --timeout 60 to increase per-response timeout.

Too verbose

Use --max-turns 2-3 for quick comparisons.