Preview

Preview Feature — This feature is currently in preview and under active development. APIs and functionality may change. We recommend testing thoroughly before using in production.

Agents

Agents are autonomous AI systems that can reason, plan, and execute multi-step tasks using tools. Unlike simple LLM calls, agents maintain context, use tools to interact with external systems, and make decisions about what actions to take.

When to Use Agents

Use Agents when you need:

Multi-step reasoning - Tasks that require planning and multiple actions
Tool usage - AI that can call APIs, query databases, or interact with systems
Autonomous decisions - Systems that decide what to do next based on context
Human oversight - AI with approval gates for sensitive operations
Persistent memory - Conversations that remember past interactions

Quick Example

Here's a simple customer support agent:

TypeScript

import Ductape from '@ductape/sdk';

const ductape = new Ductape({
  accessKey: 'your-access-key',
});

// Define the agent
const agent = await ductape.agents.define({
  tag: 'support-agent',
  name: 'Customer Support Agent',
  model: {
    provider: 'anthropic',
    model: 'claude-sonnet-4-20250514',
    temperature: 0.7,
  },
  systemPrompt: `You are a helpful customer support agent.
You have access to tools to look up orders and customer information.
Always be polite and helpful.`,
  tools: [
    {
      tag: 'lookup-order',
      description: 'Look up order details by order ID',
      parameters: {
        orderId: {
          type: 'string',
          description: 'The order ID to look up',
          required: true,
        },
      },
      handler: async (ctx, params) => {
        return ctx.database.query({
          database: 'orders-db',
          event: 'find-order',
          params: { id: params.orderId },
        });
      },
    },
    {
      tag: 'get-customer',
      description: 'Get customer information',
      parameters: {
        customerId: {
          type: 'string',
          description: 'The customer ID',
          required: true,
        },
      },
      handler: async (ctx, params) => {
        return ctx.database.query({
          database: 'customers-db',
          event: 'find-customer',
          params: { id: params.customerId },
        });
      },
    },
  ],
});

// Run the agent
const result = await ductape.agents.run({
  tag: 'support-agent',
  input: 'Can you check the status of order ORD-12345?',
});

console.log('Response:', result.output);
console.log('Tools used:', result.toolCalls.length);

How Agents Work

Agents follow a ReAct (Reason + Act) pattern:

User Input → OBSERVE → THINK → ACT → EVALUATE → Output
                ↑                        ↓
                └────────────────────────┘
                     (loop until done)

OBSERVE - Gather context from input, memory, and tools
THINK - LLM reasons about what to do next
ACT - Execute tool calls or generate response
EVALUATE - Check if task is complete or needs more steps

Agent Loop

TypeScript

while (!done) {
  // 1. Build context
  const context = buildContext(input, memory, toolResults);

  // 2. Call LLM
  const response = await llm.generate(context, tools);

  // 3. Execute tools if requested
  if (response.toolCalls) {
    for (const call of response.toolCalls) {
      const result = await executeTool(call);
      toolResults.push(result);
    }
  }

  // 4. Check termination conditions
  done = response.stopReason === 'end_turn' ||
         iterations >= maxIterations ||
         tokenUsage >= maxTokens;
}

Core Components

Model Configuration

TypeScript

model: {
  provider: 'anthropic',       // LLM provider
  model: 'claude-sonnet-4-20250514',  // Model name
  temperature: 0.7,            // Creativity (0-2)
  maxTokens: 4096,             // Max response tokens
  topP: 0.9,                   // Nucleus sampling
  stopSequences: ['DONE'],     // Custom stop sequences
  timeout: 30000,              // Request timeout (ms)
}

Supported Providers

Provider	Models	Status
`anthropic`	Claude 3.5 Sonnet, Claude 3 Opus/Sonnet/Haiku	Supported
`openai`	GPT-4, GPT-4 Turbo, GPT-3.5 Turbo	Supported
`google`	Gemini Pro, Gemini Ultra	Coming Soon
`cohere`	Command R, Command R+	Coming Soon

Tools

Tools are functions the agent can call:

TypeScript

tools: [
  {
    tag: 'search-products',
    name: 'Search Products',  // Optional display name
    description: 'Search for products by name or category',
    parameters: {
      query: {
        type: 'string',
        description: 'Search query',
        required: true,
      },
      category: {
        type: 'string',
        description: 'Product category',
        enum: ['electronics', 'clothing', 'home'],
      },
      limit: {
        type: 'number',
        description: 'Max results',
        default: 10,
      },
    },
    requiresConfirmation: false,  // Needs human approval?
    timeout: 5000,                // Tool timeout (ms)
    retries: 2,                   // Retry on failure
    handler: async (ctx, params) => {
      // Tool implementation
      return results;
    },
  },
]

Memory

Configure short-term and long-term memory:

TypeScript

memory: {
  shortTerm: {
    maxMessages: 50,              // Conversation history limit
    truncationStrategy: 'summarize',  // 'fifo' | 'sliding_window' | 'summarize'
    includeToolResults: true,     // Include tool outputs in history
  },
  longTerm: {
    enabled: true,
    vectorStore: 'agent-memory',  // Vector config tag
    retrieveTopK: 5,              // Memories to retrieve
    minSimilarity: 0.7,           // Minimum relevance
    autoStore: true,              // Auto-store interactions
    namespace: 'support-agent',   // Memory isolation
  },
}

Termination

Control when the agent stops:

TypeScript

termination: {
  maxIterations: 25,    // Max reasoning loops
  maxTokens: 50000,     // Max total tokens
  timeout: 300000,      // Max execution time (5 min)
  allowManualStop: true,  // Allow external stop signal
}

Human-in-the-Loop

Add approval gates for sensitive operations:

TypeScript

humanInLoop: {
  enabled: true,
  alwaysRequireApproval: ['refund-order', 'delete-account'],
  approvalTimeout: 300000,  // 5 minutes
  approvalWebhook: 'https://your-app.com/approve',
}

Streaming

Stream agent events in real-time:

TypeScript

streaming: {
  enabled: true,
  events: ['thinking_start', 'message_delta', 'tool_start', 'tool_end'],
  webhook: 'https://your-app.com/stream',
}

Agent Lifecycle

1. Define

Create an agent definition:

TypeScript

const agent = await ductape.agents.define({
  tag: 'my-agent',
  name: 'My Agent',
  model: { provider: 'anthropic', model: 'claude-sonnet-4-20250514' },
  systemPrompt: 'You are a helpful assistant.',
  tools: [...],
});

2. Register (Optional)

TypeScript

await ductape.agents.register('my-product', agent);

3. Run

Execute the agent:

TypeScript

const result = await ductape.agents.run({
  tag: 'my-agent',
  input: 'User message',
  sessionId: 'session-123',  // For memory continuity
});

4. Control

Send signals to running agents:

TypeScript

// Stop execution
await ductape.agents.signal({
  executionId: result.executionId,
  signal: 'stop',
});

// Approve tool call
await ductape.agents.signal({
  executionId: result.executionId,
  signal: 'approve',
  payload: { requestId: 'req-123' },
});

Execution Results

TypeScript

const result = await ductape.agents.run({...});

console.log(result.status);         // 'completed' | 'failed' | 'stopped'
console.log(result.output);         // Final response text
console.log(result.iterations);     // Number of reasoning loops
console.log(result.toolCalls);      // List of tool calls made
console.log(result.tokenUsage);     // Token usage breakdown
console.log(result.executionTime);  // Total time (ms)
console.log(result.conversationHistory);  // Full conversation

Platform Integration

Agents can access all Ductape components through the tool context:

Component	Access	Description
Actions	`ctx.api.run()`	Call external APIs
Databases	`ctx.database.query()`	Database operations
Graphs	`ctx.graph.query()`	Graph database queries
Storage	`ctx.storage.upload()`	File operations
Notifications	`ctx.notification.email()`	Send notifications
Publishing	`ctx.messaging.produce()`	Message broker publishing
Vectors	`ctx.recall()` / `ctx.remember()`	Vector memory
Features	`ctx.feature.run()`	Run features

Next Steps

Getting Started - Build your first agent
Tools - Creating and managing tools
Memory - Configure agent memory
Human-in-the-Loop - Add approval gates
Examples - Real-world agent patterns

When to Use Agents​

Quick Example​

How Agents Work​

Agent Loop​

Core Components​

Model Configuration​

Supported Providers​

Tools​

Memory​

Termination​

Human-in-the-Loop​

Streaming​

Agent Lifecycle​

1. Define​

2. Register (Optional)​

3. Run​

4. Control​

Execution Results​

Platform Integration​

Next Steps​

See Also​