Multi-Agent Routing

Triggerfish supports routing different channels, accounts, or contacts to separate isolated agents, each with its own workspace, sessions, personality, and classification ceiling.

Why Multiple Agents?

A single agent with a single personality is not always enough. You may want:

• A personal assistant on WhatsApp that handles calendar, reminders, and family messages.
• A work assistant on Slack that manages Jira tickets, GitHub PRs, and code reviews.
• A support agent on Discord that answers community questions with a different tone and limited access.

Multi-agent routing lets you run all of these simultaneously from a single Triggerfish installation.

How It Works

  Inbound Messages                      Agent Workspaces

  WhatsApp         ----------------->   Agent: "Personal"
  (personal #)                          Skills: calendar, email
                                        Classification: PERSONAL

  Slack            ----------------->   Agent: "Work"
  (work workspace)                      Skills: jira, github
                                        Classification: CONFIDENTIAL

  Telegram         ----------------->   Agent: "Side Project"
  (@projectbot)                         Skills: coding, deploy
                                        Classification: INTERNAL

The AgentRouter examines each inbound message and maps it to an agent based on configurable routing rules. If no rule matches, messages go to a default agent.

Routing Rules

Messages can be routed by:

Criteria	Description	Example
Channel	Route by messaging platform	All Slack messages go to "Work"
Account	Route by specific account within a channel	Work email vs personal email
Contact	Route by sender/peer identity	Messages from your manager go to "Work"
Default	Fallback when no rule matches	Everything else goes to "Personal"

Configuration

Define agents and routing in triggerfish.yaml:

agents:
  list:
    - id: personal
      name: "Personal Assistant"
      channels: [whatsapp-personal, telegram-dm]
      tools:
        profile: "full"
      model: claude-opus-4-5
      classification_ceiling: PERSONAL

    - id: work
      name: "Work Assistant"
      channels: [slack-work, email-work]
      tools:
        profile: "coding"
        allow: [browser, github]
      model: claude-sonnet-4-5
      classification_ceiling: CONFIDENTIAL

    - id: support
      name: "Customer Support"
      channels: [discord-server]
      tools:
        profile: "messaging"
      model: claude-haiku-4-5
      classification_ceiling: PUBLIC

Each agent specifies:

• id -- Unique identifier for routing.
• name -- Human-readable name.
• channels -- Which channel instances this agent handles.
• tools -- Tool profile and explicit allow/deny lists.
• model -- Which LLM model to use (can differ per agent).
• classification_ceiling -- Maximum classification level this agent can reach.

Agent Identity

Each agent has its own SPINE.md defining its personality, mission, and boundaries. SPINE.md files live in the agent's workspace directory:

~/.triggerfish/
  workspace/
    personal/
      SPINE.md          # Personal assistant personality
    work/
      SPINE.md          # Work assistant personality
    support/
      SPINE.md          # Support bot personality

Isolation

Multi-agent routing enforces strict isolation between agents:

Aspect	Isolation
Sessions	Each agent has independent session space. Sessions are never shared.
Taint	Taint is tracked per-agent, not across agents. Work taint does not affect personal sessions.
Skills	Skills are loaded per-workspace. A work skill is not available to the personal agent.
Secrets	Credentials are isolated per-agent. The support agent cannot access work API keys.
Workspaces	Each agent has its own filesystem workspace for code execution.

WARNING

Inter-agent communication is possible through sessions_send but is gated by the policy layer. One agent cannot silently access another agent's data or sessions without explicit policy rules allowing it.

Default Agent

When no routing rule matches an inbound message, it goes to the default agent. You can set this in configuration:

agents:
  default: personal

If no default is configured, the first agent in the list is used as the default.