Skip to content
voxlang.orgv0.6.0SSOT

Journey: Building Resilient AI Agents

The Broken Reality of Orchestrating LLMs

Section titled “The Broken Reality of Orchestrating LLMs”

Building an intelligent AI agent generally involves duct-taping language models to your application state. This requires writing brittle Python scripts or complex TypeScript orchestrators like Langchain.

As soon as your agent needs to execute a tool reliably, parse JSON tool-call responses, retry failures, and maintain a stateful memory of the interaction, the infrastructure complexity explodes. LLMs hallucinate arguments, drop nested fields, and break your application logic.

The Vox Paradigm: Built-In, Type-Safe Orchestration

Section titled “The Vox Paradigm: Built-In, Type-Safe Orchestration”

Vox was explicitly designed as an AI-native programming language. You do not need an external orchestration library to build an agent, because Vox natively generates Model Context Protocol (MCP) tool schemas and natively coordinates stateful LLM queries.

In Vox, the chaos of generative models is bounded by the compiler’s zero-null guarantees (Result and Option). You define the rigid boundaries; Vox handles the plumbing.

Core Snippet: Creating an Agent Tool

Section titled “Core Snippet: Creating an Agent Tool”

By declaring a bare tool, Vox parses the docstring, the types, and the return structure, turning your function into a ready-to-execute schema for your LLM.

type SearchResult =
    | Found(text: str, score: int)
    | NotFound(query: str)


tool "search_knowledge: Search the knowledge base for documents matching the query" search_knowledge(query: str, max_results: int) to SearchResult {
    if max_results <= 0 {
        return NotFound(query)
    }
    return Found("Document matching " + query, 95)
}


server get_answer(user_question: str) to Result[str] {
    let hit = search_knowledge(user_question, 5)
    return match hit {
        Found(text, _score) => Ok(text)
        NotFound(q) => Error("No documents found for: " + q)
    }
}

Running the Process

Section titled “Running the Process”

  1. Save the above snippet into an entrypoint like src/agent.vox.

  2. Compile and run:

    Terminal window
    vox build src/agent.vox
    vox run src/agent.vox

  3. Vox will start the development server. The endpoints become immediately queryable, and if running in MCP mode, your agent tools are automatically broadcasted for discovery.

Maturity and limitations

Section titled “Maturity and limitations”
  • Maturity: beta for tool-shaped examples — compiler and MCP registry paths evolve; treat snippets as orientation, not a guarantee every field matches shipped schemas.
  • Limitation ids: L-001 (docs may oversell partial MCP surfaces), L-023 (MCP tool registry parity is ongoing maintenance).

Deep Dives

Section titled “Deep Dives”

To truly scale out this pattern, see how Vox implements AI orchestration under the hood: