Core Module

AI Agent System

Autonomous agents that can reason, use tools, and execute multi-step tasks with full observability. Supports multi-agent teams, agent delegation, and direct playground testing.

What is an Agent?

An agent is an AI-powered execution unit that receives a goal, reasons about how to achieve it, selects and uses appropriate tools, and produces structured outputs. Agents are configured with a provider, model, role, system instructions, and a set of capabilities.

Reasoning Engine

Agents use LLMs to plan, make decisions, and adapt to unexpected situations.

Tool Selection

Automatically chooses the right tool for each sub-task from a registry of capabilities.

Memory & Context

Maintains conversation history and retrieves relevant past interactions using semantic memory.

Structured Output

Returns validated JSON responses that can be consumed by downstream steps.

Agent Teams

Group multiple agents into teams with shared objectives, war room chat, and A2A communication.

Agent Configuration

Agents are created via the Agents page or the API. Each agent requires a name, provider, and model. Optional fields include role, objective, system instructions, avatar, type, capabilities, and quota.

agent-definition.json

{
  "name": "Research Assistant",
  "description": "Summarizes and analyzes information",
  "type": "custom",
  "provider": "openai",
  "model": "gpt-4o-mini",
  "role": "researcher",
  "objective": "Find and summarize latest AI papers",
  "systemInstructions": "You are a research assistant...",
  "capabilities": ["llm", "web_search"],
  "isActive": true
}

Agent Execution Flow

Receive Goal

Agent gets a high-level task description.

Plan Actions

LLM breaks down the goal into executable steps.

Execute Tools

Agent calls tools sequentially or in parallel.

Return Result

Final output is validated and logged.

Agent Safety & Guardrails

• Maximum iteration limits prevent infinite loops.
• Tool permissions are enforced at the executor level.
• Output validation ensures structured responses.
• All actions are logged for audit and debugging.

Agent Playground

Test agents directly from the UI without creating a workflow. The playground sends prompts through the same LLM adapter used by workflow steps, so results are representative of production execution.

Agent API

POST/api/agents

Create a new agent profile.

GET/api/agents

List all registered agent profiles for the user.

POST/api/agents/:id/run

Send a prompt directly to an agent. Used by the playground.