Runtime Architecture

Overview

The Agent Runtime serves as the central coordination system for autonomous AI agents, implementing sophisticated orchestration that manages distributed AI services, character consistency, and multi-platform operations. The runtime has evolved from a simple message router to a comprehensive workflow orchestration engine supporting multi-stage LLM pipelines, tool execution, and intelligent context management.

Current Architecture

The runtime now implements a Plan & Execute pattern with the following stages:

1. Planning Stage: Analyzes user intent and determines tool usage

2. Context Enrichment: Gathers relevant context from multiple sources

3. Tool Execution: Runs identified tools in parallel when possible

4. Response Generation: Creates character-appropriate responses

5. Fallback Handling: Manages safety refusals and prompt leakage

Core Components

AgentRuntime Class

The main orchestrator that manages:

• Plugin lifecycle (initialize/shutdown)

• Message routing to appropriate handlers

• Workflow execution via WorkflowManager

• Context management and tool registration

• Conversation history per user/platform

• Error handling and fallback responses

WorkflowManager

Orchestrates multi-stage LLM pipelines:

• Executes workflow configurations (default, minimal, reasoning)

• Manages stage transitions and state

• Handles tool execution based on planning

• Provides grouped logging for UI visualization

• Supports platform-specific overrides

ContextManager

Manages context across platforms:

• Registers platform-specific context managers

• Aggregates context from multiple sources

• Provides unified context interface

• Supports Twitter timeline, news (planned), market data (planned)

ToolManager

Extensible tool framework:

• Registers and manages available tools

• Executes tools based on planning decisions

• Currently supports: web search, news, crypto prices, stock data, time, Twitter timeline

• Handles tool errors gracefully

Message Flow

Legacy Mode (Simple)

Message → Runtime → Plugin → LLM → Response

Workflow Mode (Advanced)

Message → Runtime → WorkflowManager → Planning → Context → Tools → Response → Fallbacks

Key Features

1. Multi-Stage Pipeline

// Example workflow execution
const workflow = {
  pipeline: {
    planning: { enabled: true, model: LLM.GEMINI_2_5_FLASH },
    context: { enabled: true, model: LLM.GEMINI_2_5_PRO },
    main: { model: LLM.CLAUDE_4_SONNET },
    fallbacks: {
      safety: { model: LLM.GEMINI_2_5_PRO },
      leakage: { model: LLM.CLAUDE_3_7_SONNET }
    }
  }
};

2. Tool Integration

// Tools are automatically discovered during planning
"What's the TSLA price?" → Planning identifies stock tool → Executes → Includes in response
"Latest Fed news" → Planning identifies news tool → Fetches articles → Contextual response

3. Platform-Specific Context

// Each platform can have its own context manager
runtime.contextManager.registerPlatformManager('twitter', twitterContextManager);
runtime.contextManager.registerPlatformManager('telegram', telegramContextManager);

4. Conversation Memory

// Maintains last 10 messages per user/platform
const history = runtime.getConversationHistory('twitter', 'user123');

Usage Example

// Initialize runtime with character and workflow
const runtime = new AgentRuntime(billCharacter, {
  useWorkflow: true,
  workflowConfig: defaultWorkflow
});

// Register plugins
runtime.registerPlugin(new TwitterPlugin());
runtime.registerPlugin(new TelegramPlugin());

// Start runtime
await runtime.start();

// Process messages
const response = await runtime.processMessage({
  id: '123',
  platform: 'twitter',
  content: 'What is the current market situation?',
  userId: 'user123',
  timestamp: new Date(),
  metadata: { type: 'mention' }
});

Workflow Configurations

Default Workflow

• Full pipeline with planning, context, and fallbacks

• Best for complex queries requiring tool usage

• Optimized for quality over speed

Minimal Workflow

• Direct response generation without planning

• Fastest response time

• Best for simple conversational responses

Reasoning Workflow

• Enhanced planning with reasoning tokens

• Deep analysis for complex tasks

• Supports thinking models (future)

Platform Integration

The runtime provides a unified interface for all platforms:

1. Twitter: Autonomous posting, mention responses, timeline monitoring

2. Telegram: Private chats, group conversations, channel posts

3. Website: Real-time chat, WebSocket communication

4. API: Direct programmatic access

Error Handling

The runtime implements multiple layers of error handling:

1. Plugin Failures: Fallback to default LLM response

2. Service Outages: Graceful degradation with error messages

3. Tool Failures: Continue with other tools, include error context

4. Character Consistency: All errors maintain BILL's personality

Performance Optimizations

• Parallel Tool Execution: Tools run concurrently when possible

• Context Caching: Platform context cached for 30 minutes

• Conversation Limits: Maximum 10 messages per conversation

• Async Processing: All operations are non-blocking

Future Enhancements

1. Enhanced Memory: Persistent conversation storage

2. Cross-Platform Context: Unified user profiles across platforms

3. Advanced Planning: Multi-step task decomposition

4. Streaming Responses: Real-time token streaming

5. Multi-Agent Support: Coordination between multiple characters