Building an AI Agent with TypeScript

Core Architecture

The article uses a body-system analogy: the LLM serves as the brain, the agent as the body, tools as limbs, and conversation context as a notebook tracking message history.

Key components:

• Agent class - Wraps the LLM provider (LM Studio or Anthropic) and manages the agentic loop
• Conversation class - Maintains message history with system, user, and assistant roles
• Tool interface - Standardized structure with name, description, parameters, and async execute method

The Agentic Loop

The system runs iteratively:

1. Send prompt to the LLM
2. Check response for tool call requests
3. Execute requested tools
4. Reinject results into conversation history
5. Repeat until the LLM produces a final response

This pattern enables autonomous multi-step reasoning without manual intervention.

Implementation Details

The author uses qwen2.5-7b-instruct with LM Studio for local inference on Apple Silicon. The architecture remains provider-agnostic—swapping to Anthropic requires only changing the agent class.

Example tools include a WeatherTool and ActivityTool, demonstrating how agents combine multiple tool calls to answer complex queries like "What should I do today given the weather?"

Practical Notes

• TypeScript with strict configuration catches type errors early
• Bun serves as the runtime
• The interactive REPL uses Node's readline module
• Code will age as SDKs evolve, but the underlying patterns remain stable

Connections

Implements concepts from agentic-design-patterns, particularly the tool-use and ReAct patterns. Complements build-autonomous-agents-using-prompt-chaining-with-ai-primitives with a TypeScript-specific implementation. The emphasis on provider abstraction aligns with 12-factor-agents principles.