Ant Group Open-Sources Avernet to Organize Human-Agent Collaboration

Lead

AI coding tools and agent systems have accelerated the generation of content, code and solutions. But in enterprise environments, the harder problem is no longer just making one agent smarter. It is how to discover, connect and coordinate capabilities that are scattered across teams, platforms and permission boundaries.

Ant Group has open-sourced Avernet, a multi-agent collaboration infrastructure released under the Apache 2.0 license. Its first community version focuses on the agent collaboration network, aiming to help humans and heterogeneous agents work more like an organization.

Key points

• A collaboration infrastructure for multiple agents
  Avernet is designed for human participants and different types of agents. Instead of assuming that all data, permissions and tasks can be handed to one “super agent,” it provides a framework where participants collaborate within defined boundaries.

• The community edition starts with the collaboration network
  The first open-source release supports agent discovery, joining, invitation, task participation and result return. Agents can actively join the network or connect from existing platforms, which may reduce the friction of integrating diverse agent systems.

• Four enterprise collaboration pain points
  According to the source material, Avernet originated from Ant Group’s internal collaboration needs. It targets four problems: difficulty finding the right agent, difficulty aligning multiple participants, slow task execution caused by manual handoffs, and the loss of project experience that could otherwise become organizational capability.

• Governance is a core design concern
  In real companies, data privacy, compliance and business boundaries limit what agents can access or do. The community edition has partially opened capabilities including identity authentication, access authorization, permission control, lifecycle management and some security protections. These functions address basic questions such as who an agent is, what it can see and what it can do.

• Not tied to one model or agent engine
  Avernet does not bind developers to a single model or agent runtime. Teams can connect and combine existing capabilities according to their own business scenarios, which is important for enterprises that already operate diverse AI stacks.

Why it matters

The significance of Avernet is that it frames multi-agent systems as an organizational problem, not merely a tool orchestration problem. In production environments, agents may live inside customer service, engineering, operations, risk control or other systems. Without discovery, authorization and traceability, it is difficult for them to participate safely in complex workflows.

The source states that, as of July 31, 2026, Avernet-related capabilities had been deployed across 12 core business sectors inside Ant Group, with agent task completion rates stably above 90%. While the public release is still a community edition, this background suggests that the project is based on internal production experience rather than a pure concept demo.

Future versions are expected to further open audit tracing, observability and evaluation, memory and continuous optimization, as well as service deployment and container cluster management. If these capabilities mature, Avernet could become a foundation for enterprises building multi-agent systems: not a replacement for models, but a layer that organizes models, tools, people and processes.

For developers, the open-source release lowers the engineering barrier to multi-agent collaboration. For enterprises, it points to a practical route for agent adoption: define identity, permissions and collaboration rules first, then scale automation.

Source: QbitAI