inclusionAI/Avernet v2026.09.17

Avernet v2026.09.17

Repository: inclusionAI/Avernet

Tag: v2026.09.17

Published: 2026-10-05T08:15:02Z

Prerelease: no

Release notes:

Avernet v2026.09.17

> Release branch: REL20260917 · previous release: v2026.08.04 · 717 pull requests · 42 contributors · 中文版见文末

Overview

Avernet is an open-source infrastructure layer for building and operating persistent, coordinated, multi-agent systems at organizational scale. v2026.09.17 is the first official release since v2026.08.04. It merges 717 pull requests and moves forward on three fronts:

• Organizational scale: Spaces become the boundary for team collaboration and governance; with the Bot owner's authorization, an application can make calls on a user's behalf; Bot configuration can be managed and audited through a single declarative manifest.
• Multi-agent collaboration: Group chats get a persistent congestion-control queue; agents can ask people questions mid-run, and human members take part under their own identity; goal-driven tasks can be passed between multiple Bots until they are done.
• Ecosystem integration: DeepSeek Harness and OpenClaw join the collaboration network through BCN plugins; collaboration APIs and Bot registration are exposed uniformly through the Gateway's OpenAPI v1; the community edition can be deployed to Alibaba Cloud ACK.

This release includes several breaking API and configuration changes. Read Upgrade Notes before upgrading.

Highlights

Organizational scale: Spaces as the boundary for collaboration, authorization, and governance

• Spaces: Adds personal Spaces and team Spaces. Users join a team Space by submitting a join request; a team Space holds up to 20 Bots by default, adjustable per Space; Space members automatically get MEMBER permission on the Bots in the Space. #1448 #1497 #1860 #2295
• Skill collaboration and publishing in Spaces: Team members can maintain Skills together within a Space. Edit requests are approved through work orders, and draft edits are protected by a lease; publishing can resume after an interruption and resolves to a specific version in Skill Center. #1658 #1701 #1709
• Application authorization: A Bot owner can authorize an application to access their Bot. The application uses machine credentials to call on the grantor's behalf when no user is signed in; application API keys and authorizations are now managed in the Gateway. #937 #951 #1023 #1699
• Declarative Bot configuration manifest: A single manifest describes a Bot's Skills, identity files, resources, and MCP and CLI tools. The platform fetches and validates it, keeping every fetched byte for auditing; a manifest can be supplied when a Bot is created, and updates take effect at the matching lifecycle point for each engine. #1727 #1758 #1791 #1836 #1870
• Collaboration governance and observability: Multi-user editing now uses declarative edit locks that collaborators can acquire or take over. Adds Agent observability, which shows each Bot's health status and diagnosis history; owners and administrators can advance improvement items to acceptance. #1511 #2170 #2135
• Open-source Agent self-evolution loop: Agent self-evolution is split into a modular web runtime and ships diagnosis, optimization, and evaluation Skills that run in local OpenClaw and Singlebox; issues can be aggregated across multiple runs and fixed together by issue group. #1946 #2168 #2129 #2197

Multi-agent collaboration: reliable group chats, with humans always in the loop

• Message congestion control: BCS adds a database-backed delivery queue. Messages for the same Bot and session run in order; concurrency, send interval, and queue capacity can be limited per Bot; system context such as session initialization is queued as well. The queue is off by default and can be enabled gradually by source type and by Bot. #1970 #2199
• Recoverable execution: A delivery in an uncertain state can be settled after the sender or the Bot owner verifies it; session callbacks can be safely taken over after an instance exits, and failed state-machine runs can be rerun with their history kept; in group chats, an abort affects only the target Bot. #2181 #1645 #1762
• Human in the loop (HITL): Agents can request command approval, ask the user a question, or switch modes mid-run; once the person answers, the same run continues. #1081 #1430
• Humans as first-class participants: People can join a session in a participation mode of their choice, and can also choose to see only the messages relevant to them (useful for judge-and-player scenarios); when someone is @-mentioned, the notification can go to several channels at once and into a work order. #1102 #1986...