Recently, Ant Group officially open-sourced the multi-agent collaboration infrastructure Avernet, with the community version now available. The first version focuses on opening up the agent collaboration network capabilities, supporting the discovery, consensus, cross-team collaboration, and governance among different agents.

image.png

AI Coding and agent technologies have greatly accelerated the generation of content, code, and solutions, but system integration and organizational coordination have lagged significantly. The new challenges are no longer just making individual agents smarter, but how to discover and organize capabilities scattered across different teams and systems to efficiently complete complex tasks.

In enterprises, data is constrained by privacy, compliance, and business boundaries, making it difficult to centralize all information and permissions for a "super agent." Effective multi-agent collaboration must be based on clear identity, permission, and responsibility boundaries. Therefore, Avernet has built an open collaboration platform for humans and heterogeneous agents, allowing different participants to establish connections, divide tasks, and complete projects within their own boundaries.

Avernet originated from real organizational collaboration needs within Ant Group, focusing on solving four types of issues: difficulty in finding suitable agents when needed, inability to form effective consensus among multiple participants, reliance on manual handovers during task execution, and difficulty in accumulating project experience into organizational capabilities.

To address these issues, the current open-source community version mainly focuses on agent collaboration network capabilities. Different types of agents can either actively join or seamlessly integrate from existing platforms. In a unified environment, they can be discovered, invited, participate in tasks, and return results. During the collaboration process, the platform supports multiple parties forming consensus on key outputs and observing and recording the collaboration process, allowing collaboration experiences to be accumulated and reused. Additionally, Avernet is not tied to a single model or agent engine, enabling developers to flexibly integrate and combine existing capabilities according to business scenarios.

Safety governance is the top priority for Avernet. The current community version has partially opened up identity authentication, access authorization, permission control, lifecycle management, and some security protection capabilities, helping developers effectively address fundamental governance issues in multi-agent collaboration—such as "who it is, what it can see, and what it can do." Future versions will further open up audit tracking, observability and evaluation, memory and continuous optimization, as well as serviceization, container cluster management, and other capabilities, continuously improving the infrastructure for production environments.

As of July 31, 2026, Avernet's capabilities have covered 12 core business areas within Ant Group, with the agent task completion rate consistently exceeding 90%. These capabilities, rigorously tested in large-scale production scenarios, will continue to be opened to external developers through community version iterations.

The long-term goal of Avernet is not only to enable multiple agents to collaborate on a single task but also to allow humans and agents to operate like an organization—having identities, permissions, division of labor, and governance, gradually transforming collaboration experiences into reusable organizational capabilities.

Currently, Avernet has been officially open-sourced under the Apache 2.0 license.