Open Agent Loops

Open Agent Loops is a minimal, provider-agnostic framework designed for building agent loops, focusing on modularity and flexibility. The tool is structured around swappable interfaces for key components such as models, memory, tools, and stop conditions, allowing developers to substitute implementations without changing the core logic. It operates headlessly by default, emitting typed event streams rather than rendering output, so users can integrate their own front ends or consume the data in various formats including CLI, DOM timelines, or raw JSONL. The framework exposes a concise API surface, centering on the runAgent function. Developers can define tools—each with a name, description, parameter schema, and execution function—and pass them into the agent loop. The memory interface is also modular, with implementations such as SessionMemoryStore available and the ability to swap to alternatives like Redis. The model interface is compatible with any endpoint that supports the OpenAI chat-completions wire format, and the evidence lists open-model families such as DeepSeek, GLM, Qwen, Kimi, MiniMax, and Gemma as being exercised with the framework. Tool calls can run in parallel, and results are folded back into the loop. Stop conditions are configurable, allowing the agent to halt on criteria such as a maximum number of steps or a custom predicate. Streaming is a core feature, with the stream() method returning an async iterable of events, including reasoning, text, and tool calls, delivered incrementally. The tool supports optional hooks for extending behavior, such as gating tool calls or reshaping context. Additional composable building blocks are mentioned, including skills (bundled instructions and tools), planning tools (durable working memory and workflow freezing), composable agents (sub-agents callable as tools), and channels for integrating live transports like Slack or Discord. Open Agent Loops is delivered as a universal ESM package with a single dependency (zod), running in environments such as Node, Bun, Deno, and browsers. The evidence does not specify pricing or licensing details.