Deuz SDK v2.0.0: TypeScript Framework for Production AI Agents

What changed

The Deuz SDK, a TypeScript framework designed for production AI agents, has seen its latest release, version 2.0.0. This framework emphasizes zero runtime dependencies and offers a unified streaming API for interacting with various large language models, including Claude, GPT, Gemini, Grok, Mistral, and DeepSeek. It supports multiple execution environments such as Node.js, Bun, Deno, serverless, and edge runtimes.

Key features of the Deuz SDK include:

• Models: Support for multiple providers, streaming responses, structured output, and tool calls.
• Agents: Enable typed results, verification, human-in-the-loop approvals, and resumable agent runs.
• Swarms: Facilitate task dependencies, runtime spawning of agents, shared data blackboards, round-based execution, and persistence using SQLite or PostgreSQL.
• Operations: Includes features for leases, draining, cross-process cancellation and recovery, persistent budgets, and scheduling.
• Context: Provides memory management, retrieval, compaction, and Multi-modal Context Processing (MCP) tools.
• Control: Offers shared execution policies, budget accounting, tracing, and optional React bindings.

The SDK is built with TypeScript and is compatible with Node.js 22+ or edge runtimes with Web API support. Integrations for Node.js are available via separate imports.

Why it matters for builders

This SDK provides a comprehensive toolkit for developers looking to build complex AI agents. The zero-dependency nature simplifies integration, while the unified API for multiple LLMs reduces vendor lock-in and allows for easier model switching. Features like durable execution and human-in-the-loop approvals are critical for production-grade agent deployments.

Practical impact

Developers can leverage the Deuz SDK to create agents with advanced capabilities such as long-term memory, planning, and sandboxed code execution (CodeAct). The framework's support for hybrid RAG (Retrieval-Augmented Generation) and tool calling enhances agent intelligence and utility. The ability to manage agent swarms and their interdependencies offers a path to building more complex, multi-agent systems. Builders can start by installing the core SDK via npm and exploring the provided examples for running their first agent.

Caveats and source limits

The provided source is primarily a GitHub repository description and excerpt. Specific benchmark results, performance metrics, or detailed comparisons to other agent frameworks are not available. While the release mentions version 2.0.0, specific details about what changed between previous versions and 2.0.0 are limited to the general feature list. The source also indicates that without a lease provider, swarm runs must be driven from a single process, and external effects may require reconciliation before retries.