04-worker-mode

Worker Mode

Worker mode allows an agent to dispatch complex, long-running tasks to background "copy" agents while the main agent stays fully interactive. When a worker finishes, its result is automatically surfaced back to the user through the main agent's conversation.

This mode is ideal when users submit tasks that take a while (research, code generation, multi-step analysis) and you want them to continue chatting or submit more tasks without blocking.

See the Worker Mode Example for a complete, runnable implementation.

Enabling Worker Mode

import { Agent, TAgentMode } from '@smythos/sdk';

const agent = new Agent({
    name: 'Research Assistant',
    behavior: 'You are a research assistant capable of handling complex tasks.',
    model: 'gpt-4o',
    mode: TAgentMode.WORKER,
});

How It Works

Task Dispatching

When the user asks for something complex, the main agent's LLM evaluates the request:

• Simple tasks (quick questions, math, lookups): The agent answers directly — no worker needed.
• Complex tasks (research, analysis, code generation): The agent dispatches the task to a background worker via the internal _sre_Worker_Dispatch skill.

The main agent immediately confirms the dispatch and remains available for new prompts.

Copy Agents

Each worker is a "copy" of the main agent — it shares the same skills and model but runs independently in its own chat session. The copy agent:

• Preserves the original agent behavior (system prompt) defined by the developer
• Gets all of the main agent's skills (except internal _sre_* skills)
• Receives additional instructions appended to the behavior, telling it to use <worker_result> and <worker_question> tags for communication
• Has no awareness of the main conversation — it only knows its assigned task

Worker Communication

Workers communicate back to the main agent through two tag-based protocols: