Eko: Open‑Source Agent Framework for Building Production‑Ready Virtual Employees

The article introduces Eko, an open‑source agent development framework created by researchers from Tsinghua, Fudan, Stanford and others, designed to let developers build production‑ready "virtual employees" using concise code and natural language.

Key innovations include:

Mixed Agentic Representation: Seamlessly blends high‑level natural‑language design with low‑level programming language implementation.

Cross‑Platform Agent Framework: A single codebase runs in browsers, as browser extensions, or in Node.js environments, thanks to an environment‑aware architecture.

Production‑Level Intervention Mechanism: Explicit hooks allow human oversight and real‑time adjustments, ensuring safe governance of AI‑driven workflows.

The environment‑aware architecture consists of three layers: a universal core providing workflow management and LLM integration, environment‑specific tools (e.g., browser extensions, web, Node.js), and an environment bridge that handles detection, registration, resource management, and security.

Security measures are tailored per environment: strict permission controls for browsers, broader system access for Node.js with user‑prompted confirmations.

Automatic tool registration (e.g., loadTools() ) enables seamless switching across environments.

Hierarchical planning separates tasks into a Planning layer that converts user intents into a domain‑specific task graph, and an Execution layer that synthesizes concrete actions and tool calls, with automatic merging of consecutive LLM calls to speed inference.

The Visual‑Interactive Element Perception (VIEP) framework improves browser automation by combining visual recognition with element context, generating compact pseudo‑HTML representations that dramatically reduce processing overhead.

Eko’s hook system offers three levels—Workflow Hooks, Subtask Hooks, and Tool Hooks—allowing developers to inject custom logic before/after workflows, sub‑tasks, or individual tool executions, supporting monitoring, validation, error handling, and even human approvals.

Resources: Homepage, GitHub repository, and documentation links are provided for further exploration.