Build AI Teams Like Lego: Introducing OxyGent for Scalable Collective Intelligence

Problem

Existing multi‑agent development is hampered by ecosystem fragmentation, cumbersome configuration, and opaque decision logic, which makes building and maintaining systems costly for individual developers and small teams.

Design Principles

OxyGent treats tools, models, and agents as plug‑in atomic operators named Oxy . Each Oxy component is hot‑pluggable and can be assembled directly in pure Python without YAML files. The framework provides a closed‑loop pipeline that connects three stages: build → inference → evolution .

Core Capabilities

Rapid assembly : developers create a multi‑agent system in minutes by importing Oxy components and wiring them together in code, similar to snapping Lego bricks.

Dynamic collaboration : agents self‑organize in a group‑chat‑like conversation, negotiate roles, and invoke APIs without a rigid workflow.

Full‑traceability : every inference step, tool call, and negotiation is recorded in a Git‑style decision graph that can be inspected, paused, or rewound.

Fault injection & parallel experiments : the runtime allows pausing execution, modifying an agent’s internal state (e.g., prompts, memory snapshots, pending tools), injecting failures, and running multiple model variants concurrently; all variants are automatically logged.

Continuous evolution : a built‑in evaluation engine can freeze and replay agent states, supporting supervised‑fine‑tuning (SFT) and reinforcement‑learning data collection for iterative improvement.

Elastic architecture : a distributed scheduler enables linear scaling according to Metcalfe’s law, allowing arbitrary agent topologies from simple ReAct loops to complex hybrid planners.

Performance & Benchmark

OxyGent includes production‑grade time tracing that breaks down per‑second consumption of LLM inference, tool/API calls, and agent negotiation, exposing bottlenecks instantly. On the GAIA (General AI Assistant) benchmark, OxyGent achieved a score of 59.14 , the highest among open‑source frameworks; the proprietary OWL framework scored 60.8.

Agent Lifecycle Management

Write agents in pure Python (no YAML).

Deploy with a single command locally or in the cloud.

Trace every decision step end‑to‑end via the generated decision graph.

Agents evolve automatically through continuous feedback from the evaluation engine.

Repository

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