Build Flexible Multi‑Agent Systems Like LEGO with OxyGent – New Features Unveiled

Release 1.0.8

OxyGent 1.0.8 adds multimodal information exchange, fine‑grained message control, MCP reconnection with request‑header support, and front‑end streaming output. Install the new version with pip install oxygent==1.0.8.

Design Philosophy

The framework follows a stateless, AOP‑inspired architecture. Agent, Tool, LLM and Flow are abstracted as interchangeable "Oxy" components. System composition is expressed through permission relationships between Oxy units, which enables dynamic planning and real‑time generation of execution flowcharts.

Execution Lifecycle

_pre_process – Data cleaning and formatting before any operation.

_pre_log – Log upcoming tool calls for traceability.

_pre_save_data – Persist input state prior to execution.

_format_input – Ensure input conforms to downstream requirements.

_pre_send_message – Dispatch formatted data to downstream components.

_before_execute – Load memory or other resources needed for execution.

_execute – Core computation or processing logic.

_after_execute – Post‑processing such as resource release or initial result analysis.

_post_process – Custom output handling before the next stage.

_post_log – Record results and observations for later analysis.

_post_save_data – Store final results or state.

_format_output – Adapt output format for downstream consumption.

_post_send_message – Send the final message to downstream components.

This ordered sequence guarantees deterministic data flow and supports dynamic planning with live flow‑chart generation.

Data Scopes

Four hierarchical data domains are provided:

Application – Global data accessed via oxy_request.get/set_global_data().

SessionGroup – Shared across a group of sessions via oxy_request.get/set_group_data().

Request – Per‑request data via oxy_request.get/set_shared_data().

Node – Arguments specific to a single Oxy node via oxy_request.get/set_arguments().

These APIs enable isolated yet shareable state across agents, improving data‑management efficiency in multi‑agent collaborations.

Typical Use Cases

SOP workflow – Decompose business processes into agents for approval, validation, and notification, improving traceability and efficiency.

RAG QA – Layered agents handle task allocation, knowledge retrieval, answer generation, and review, boosting relevance of generated answers.

Data analysis – Agents automate collection, cleaning, modeling, and visualization, supporting auto‑reporting and anomaly detection.

Tool invocation – Top‑level agents select external tools, while dedicated agents execute and monitor calls, enabling cross‑system automation.

Hierarchical classification – Multi‑level agents progressively refine categories and attach metadata, useful for tagging, product tiering, or risk stratification.

Community Extensions

Developers have built extensions such as:

Natural‑language‑to‑SQL conversion with automatic execution and metric validation.

Low‑code web‑scraping and file manipulation using JoyCode, driven by natural language.

Automatic generation and rendering of flow‑chart code via local API calls.

Slow‑SQL diagnosis that produces remediation suggestions.

These examples demonstrate the framework’s extensibility.

Getting Started

A concise tutorial shows how to launch a basic agent in roughly 20 lines of code. The tutorial covers environment installation, model registration, MCP tool integration, agent registration, visual debugging, multi‑agent collaboration, and distributed deployment.

Repository

GitHub: https://github.com/jd-opensource/OxyGent