Agent Examples

LLM agent workflow examples for the Loopstack automation framework.

Four ways to build agents in Loopstack — from the recommended AgentWorkflow shortcut to a from-scratch custom loop. Pick the right shape for your control vs. simplicity tradeoff.

Use in Your App

Copy this directory into your app, then install what it imports:

npm install @loopstack/agent @loopstack/claude @loopstack/code-agent @loopstack/common @loopstack/llm-provider @loopstack/mcp @loopstack/remote-client

Register the module:

import { Module } from '@nestjs/common';
import { LoopstackModule } from '@loopstack/loopstack-module';
import { AgentExamplesModule } from './agent/agent-examples.module';

@Module({
  imports: [LoopstackModule.forRoot(), AgentExamplesModule],
})
export class AppModule {}

Required app-module configuration

All four examples use LLM tools (LlmGenerateTextTool, LlmDelegateToolCallsTool, etc.) from @loopstack/llm-provider. That module is @Global and must be configured once in your root module to set the default model:

import { Module } from '@nestjs/common';
import { LlmProviderModule } from '@loopstack/llm-provider';
import { LoopstackModule } from '@loopstack/loopstack-module';
import { AgentExamplesModule } from './agent/agent-examples.module';

@Module({
  imports: [LoopstackModule.forRoot(), LlmProviderModule.forRoot({ model: 'claude-sonnet-4-6' }), AgentExamplesModule],
})
export class AppModule {}

AgentExamplesModule re-imports ClaudeModule (provider), AgentModule (the AgentWorkflow / ChatAgentWorkflow classes), CodeAgentModule, McpModule, and RemoteClientModule.forFeature({ type: 'sandbox' }). LlmProviderModule.forRoot(...) sets the default model the tools dispatch to.

The Code Agent example additionally requires RemoteClientModule.forRoot({ environments }) — see Code Agent → Setup below.

Environment

ANTHROPIC_API_KEY=sk-ant-...
LINEAR_MCP_TOKEN=...   # required only for the MCP Linear example

Examples

Example	Studio title	Description
Agent	Agent - Basic Agent Example	Launch AgentWorkflow with custom tools — recommended starting point
Code Agent	Agent - Code Agent Example	Specialized codebase-exploration agent with glob / grep / read
MCP Linear	Agent - MCP Linear Example	Chat agent connected to Linear’s hosted MCP server via mcp_list_tools + mcp_call
Custom Agent	Agent - Custom Agent Example	Build an agent loop from scratch — delegate primitives, guards, error handling

---

Agent

Launches a generic AgentWorkflow as a sub-workflow with two custom tools (weather_lookup, calculator) and renders the agent’s run inline in the parent view. The recommended starting point for agent workflows.

Files

• agent-example.workflow.ts — workflow class
• tools/weather-lookup.tool.ts, tools/calculator.tool.ts — custom tools
• templates/system.md — agent system prompt

Code Agent

Launches AgentWorkflow as a code-exploration agent with glob, grep, and read tools. Demonstrates configuring a domain-specific agent without building a custom loop.

Setup

The glob, grep, and read tools execute on a remote agent, so the workspace must have a sandbox environment connected.

1. Run a remote agent. The simplest option is @loopstack/remote-server on localhost:3001 (see registry/remote-client/server/).

2. Register an available environment in your app module:

import { RemoteClientModule } from '@loopstack/remote-client';

@Module({
  imports: [
    LoopstackModule.forRoot(),
    RemoteClientModule.forRoot({
      environments: {
        available: [
          {
            type: 'sandbox',
            name: 'Local Remote Server',
            connectionUrl: 'http://localhost:3080',
            agentUrl: 'http://localhost:3001',
            local: true,
          },
        ],
      },
    }),
    AgentExamplesModule,
  ],
})
export class AppModule {}

AgentExamplesModule itself declares the sandbox slot via RemoteClientModule.forFeature(...) — make sure you use the same type = sandbox.

3. Create a workspace from the Studio. The environment is auto-selected from the matching type and persisted on creation.

Files

• code-agent-example.workflow.ts — workflow class

MCP Linear

Launches ChatAgentWorkflow configured with mcp_list_tools + mcp_call, pointing at Linear’s hosted MCP server. The agent dynamically discovers Linear’s tool surface at runtime and invokes the right one for the user’s question.

Requires LINEAR_MCP_TOKEN env var and configures McpModule.forRoot with an allowed-hosts policy.

Files

• mcp-linear-example.workflow.ts — workflow class

Custom Agent

A from-scratch agent loop using the delegate primitives (LlmGenerateTextTool, LlmDelegateToolCallsTool, LlmUpdateToolResultTool). Demonstrates the pattern AgentWorkflow encapsulates: LLM call → guard on tool_use vs end_turn → delegate tools → feed results back → loop.

Also exercises error self-correction: the LLM is instructed to deliberately trigger validation errors (strictSchema with wrong args), runtime errors (runtimeError), and failed sub-workflows (failingSubWorkflow), then observe the is_error tool results and correct course.

When to use

• You need finer control than AgentWorkflow provides (custom guards, cancellation logic, telemetry hooks)
• You want to understand what AgentWorkflow does under the hood
• You need a custom shape — multiple parallel LLMs, conditional tool sets per turn, etc.

For the simple case, prefer AgentWorkflow (see Basic Agent Example).

Files

• custom-agent-example.workflow.ts — workflow class
• custom-agent-example.ui.yaml — Studio widget
• templates/system.md — system prompt
• tools/{strict-schema,runtime-error,failing-sub-workflow}.tool.ts — error-trigger tools
• failing-sub.workflow.ts — sub-workflow that always fails (used by failing-sub-workflow.tool)

About

Author: Jakob Klippel

License: MIT