Tool Pipelines

Tool pipelines are small, bounded workflows made from the same tools agents can already call. Use them when a user wants future communication or a recurring report that depends on agent work.

Pipelines are not a special daily-summary system. They are a primitive for composing steps such as:

• read an agent's recent output
• send that text to an existing or newly created ChatGPT+Tools or coding agent
• read the analysis agent's output later in the pipeline
• notify, text, email, call, or send a file/image to the user

What pipelines are#

Use run_tool_pipeline for immediate multi-step workflows. Use schedule_tool_pipeline when the pipeline should run later or recur.

Each pipeline is limited to a small number of declarative steps. A step can either call an existing chat.dev tool or render a safe template. Templates can reference previous step outputs with {{steps.stepId.path}}. They cannot run shell commands, Python, network requests, or arbitrary code.

Useful tools inside pipelines include:

• read_agent_output
• create_agent
• send_instructions_to_agent
• notify_user
• call_user
• send_file_to_user
• send_image_to_user
• generate_image

Pipeline steps#

Each step needs an id.

Tool step:

{
  "id": "read_builder",
  "toolName": "read_agent_output",
  "args": {
    "agentId": "builder-agent",
    "limit": 20
  }
}

Template step:

{
  "id": "summary_prompt",
  "type": "template",
  "template": "Summarize the important progress from these messages:\n\n{{steps.read_builder.messages}}"
}

Later steps can use that template output:

{
  "id": "send_to_summarizer",
  "toolName": "send_instructions_to_agent",
  "args": {
    "agentId": "daily-summary-agent",
    "instructions": "{{steps.summary_prompt.text}}"
  }
}

Scheduled reports#

For a recurring report, call schedule_tool_pipeline with a first run time and recurrence:

{
  "delaySeconds": 3600,
  "recurrence": {
    "everySeconds": 86400,
    "maxRuns": 30
  },
  "steps": []
}

The recurrence interval must be at least 60 seconds. Use maxRuns when the report should expire automatically.

Example#

This pipeline reads recent work from an agent, asks a separate ChatGPT+Tools summarizer agent to analyze it, reads the summarizer output, and emails the user.

{
  "delaySeconds": 3600,
  "recurrence": {
    "everySeconds": 86400,
    "maxRuns": 30
  },
  "steps": [
    {
      "id": "read_turtle_agent",
      "toolName": "read_agent_output",
      "args": {
        "agentId": "turtle-research",
        "limit": 30
      }
    },
    {
      "id": "build_prompt",
      "type": "template",
      "template": "You are preparing a daily user update. Summarize concrete progress, blockers, and next actions from this agent output:\n\n{{steps.read_turtle_agent.messages}}"
    },
    {
      "id": "ask_summarizer",
      "toolName": "send_instructions_to_agent",
      "args": {
        "agentId": "turtle-summary-agent",
        "instructions": "{{steps.build_prompt.text}}"
      }
    },
    {
      "id": "read_summary",
      "toolName": "read_agent_output",
      "args": {
        "agentId": "turtle-summary-agent",
        "limit": 10
      }
    },
    {
      "id": "email_user",
      "toolName": "notify_user",
      "args": {
        "channel": "email",
        "message": "Daily turtle team update:\n\n{{steps.read_summary.messages}}"
      }
    }
  ]
}

If you need a new summarizer, add a create_agent step before send_instructions_to_agent and reference the created agent ID in later steps, for example {{steps.create_summarizer.id}}.