Invect Docs

Overview

The Agent node type runs an AI agent that can iteratively call tools to accomplish a task. Unlike simple LLM calls, agents run a loop:

Send task prompt + available tools to the LLM
LLM responds with text or tool call(s)
Execute the requested tools and return results
Repeat until a stop condition is met

Agent Node Configuration

Parameter	Type	Description
`credentialId`	`string`	OpenAI or Anthropic API credential
`model`	`string`	Model ID (e.g., `gpt-4o-mini`, `claude-sonnet-4-0`)
`taskPrompt`	`string`	Main task/goal (supports Nunjucks templates)
`systemPrompt`	`string?`	Optional system instructions
`enabledTools`	`string[]`	Array of tool IDs to enable
`maxIterations`	`number`	Loop limit (1–50, default 10)
`stopCondition`	`string`	`explicit_stop`, `tool_result`, or `max_iterations`
`enableParallelTools`	`boolean`	Allow parallel tool execution

Tool Sources

Agent tools come from three places:

1. Actions (Primary)

Every action registered with defineAction() is automatically available as an agent tool. This includes all 50+ built-in actions (Gmail, Slack, GitHub, etc.) plus any custom actions you create.

2. Standalone Tools

Utility tools that don't correspond to a flow node:

math_eval — Evaluate mathematical expressions
json_logic — Apply JSON Logic rules

3. Custom Tools

You can register your own standalone tools:

custom-tool.ts

import type { AgentToolDefinition, AgentToolExecutor } from '@invect/core';

export const myToolDefinition: AgentToolDefinition = {
  id: 'my_tool',
  name: 'My Tool',
  description: 'Does something the agent needs',
  inputSchema: {
    type: 'object',
    properties: {
      query: { type: 'string', description: 'Search query' },
    },
    required: ['query'],
  },
  category: 'utility',
};

export const myToolExecutor: AgentToolExecutor = async (input, context) => {
  const result = await search(input.query);
  return { success: true, output: result };
};

Stop Conditions

Condition	Behavior
`explicit_stop`	Agent must call a "stop" function to end
`tool_result`	Stops after the first tool call returns
`max_iterations`	Stops after reaching the iteration limit

Supported Providers

Provider	Models
OpenAI	`gpt-4o`, `gpt-4o-mini`, `o1`, `o3-mini`
Anthropic	`claude-sonnet-4-0`, `claude-haiku-3-5`, `claude-opus-4`

Both providers support tool calling natively. The adapter handles converting tool definitions to the provider-specific format.

Example: Email Processing Agent

Configure an agent node with:

Task prompt: "Read the latest 5 emails from {{ gmail_query }}. Summarize each one and create a Linear issue for any that require follow-up."
Enabled tools: gmail.list_messages, gmail.get_message, linear.create_issue
Stop condition: explicit_stop
Max iterations: 20

The agent will autonomously read emails, analyze them, and create issues — calling tools as needed in a loop.

AI Agents