TAI42 — the OS for AI tools

Give a single WhatsApp number, Telegram bot, or Slack app many skills by routing it to a deep agent that triages each message to a specialist sub-agent. The bridge wiring is unchanged from the other how-tos — only the agent behind the route changes.

Why a deep agent

A conversation route runs one registered agent by name, and its turn is that agent’s astream. To fan a single conversation out to several skills, the routed agent must be able to delegate — which is exactly what a deep agent does: it plans and calls sub-agents as its skills. The plain tools_agent cannot: its input schema omits subagents and rejects them loudly (extra="forbid", plus a defense-in-depth raise on both faces). Reach for a deep agent when one bot needs many skills.

Configure the specialists

The only thing you author is the roster of specialists — each a name, a description the triage agent routes on, a system_prompt, and the tools that specialist may use. This is a DeepSubAgentSpec roster:

[
  {
    "name": "billing",
    "description": "Questions about invoices, balances, and payments.",
    "system_prompt": "You are the billing specialist. Answer only billing questions.",
    "tools": ["lookup_invoice"]
  },
  {
    "name": "tech-support",
    "description": "Troubleshooting a device or connection.",
    "system_prompt": "You are the technical-support specialist. Walk the user through fixes.",
    "tools": ["run_diagnostic"]
  }
]

Bake that roster into a small deep-agent-backed agent registered under a route-able name. The agent is a thin authored wrapper — the roster is its whole configuration:

triage_agent.py

from tai42_agents.deep_agent.agent import DeepAgent
from tai42_agents.deep_agent.tool_spec import DeepSubAgentSpec
from tai42_contract.agent import Agent
from tai42_contract.app import tai42_app

_SPECIALISTS = [
    DeepSubAgentSpec(
        name="billing",
        description="Questions about invoices, balances, and payments.",
        system_prompt="You are the billing specialist. Answer only billing questions.",
        tools=["lookup_invoice"],
    ),
    DeepSubAgentSpec(
        name="tech-support",
        description="Troubleshooting a device or connection.",
        system_prompt="You are the technical-support specialist. Walk the user through fixes.",
        tools=["run_diagnostic"],
    ),
]

_deep = DeepAgent()


@tai42_app.agents.agent("triage")
class TriageAgent(Agent):
    tool_name = "triage"
    tool_description = "Front desk that routes each message to a specialist sub-agent."
    ToolInput = DeepAgent.ToolInput

    async def run(self, *, user_message: str = "", thread_id: str | None = None, **_: object) -> str:
        return await _deep.run(user_message=user_message, thread_id=thread_id, subagents=_SPECIALISTS)

A deep agent nests one level of sub-agents — a sub-agent may not itself declare sub-agents. Each specialist’s tools are ordinary client tools the deployment has loaded, so the specialist can only use what the turn’s execution key is also allowed to run.

Load it and route to it

Name the module under agents[] — importing it registers triage — and enable the deep-agent module it builds on.

manifest.yml

agents:
  - title: tai-agents-deep
    module: tai42_agents.deep_agent
    include: [deep_agent]
  - title: triage
    module: myapp.triage_agent

Point any conversation route at triage — the channel and execution-key wiring is exactly as in the other guides:

tai conversations create support-line \
  --door channel --agent triage --execution-key support-bot \
  --channel twilio --identity whatsapp:+14155238886

Now one number answers billing and tech-support questions, each handled by its specialist, with one shared conversation thread per client.

Get started

Concepts

Guides

Marketplace

Contributing

One bot, many skills

Why a deep agent

Configure the specialists

Load it and route to it

See also

​Why a deep agent

​Configure the specialists

​Load it and route to it

​See also

Why a deep agent

Configure the specialists

Load it and route to it

See also