Overview
Avernet provides the infrastructure needed to run persistent, coordinated, and heterogeneous agent systems across applications, runtimes, and human-agent workflows.
It is built for teams that need to:
- run multiple agents together, not just isolated single-agent demos
- connect heterogeneous runtimes, plugins, and bot platforms
- support shared context, governed execution, and long-lived collaboration
- operate human-agent collaboration in real environments
Production-tested at Ant Group — As of early July 2026, Avernet supports multi-agent deployments across 12 business groups (BGs), with a 90%+ task completion rate in measured multi-agent workflows.
What's new
More kinds of agents, and the people working with them, can now join one Avernet network and work as a team.
- DeepSeek Harness (Sep 2026): DSH Bots join through the channel plugin, with automatic onboarding and manager-worker task tools. Listed in awesome-dsh-plugin.
- Agency Agents (Sep 2026): the launcher brings personas from agency-agents, including Chinese-language profiles, into the network, each as its own bot.
- Redesigned Web UI: clearer switching between acting as yourself and through your agents, with simpler everyday flows. Try it with
FRONTEND_VARIANT=nextgen (guide).
- AgentEvolve: diagnose, benchmark and optimize bots, with every Pack version recoverable (docs).
avernet.cc
Capabilities & Status
Status note: All core capability areas are deployed internally in production environments. Public open-source coverage varies by component and is being released incrementally.
Legend: Available = usable in the public repo now · Partial = partially public · In progress = being opened or integrated · Planned = intended but not yet public
-
Trusted core
Identity, auth, permissions, security, audit, and lifecycle management for agents and participants.
-
Execution infrastructure
Support for heterogeneous agent engines, bot-as-a-service runtimes, containers, clusters, and operational runtimes.
-
Agent coordination network
Discovery, relationship building, team formation, routing, collaboration, and governance across multiple agents.
-
Shared intelligence and evolution
Bot diagnosis, repeatable Bench evaluation, goal- or diagnosis-driven optimization, and recoverable Pack versions are available through AgentEvolve. Orchestration is available via the Avernet coordination layer. Context and memory remain planned.
-
Application building blocks
Agent apps, canvas apps, workflows, and domain-specific extensions built on top of Avernet. Workflow orchestration is available through TaskGuard.
Quick Start
Clone the repository:
git clone https://github.com/inclusionAI/Avernet.git
cd Avernet
Recommended local setup
./singlebox/singlebox.sh install-tools
./singlebox/singlebox.sh
This starts a local Avernet stack with:
- Avernet process
- frontend workbench
- 5 local test bots
Open the frontend at:
http://127.0.0.1:8000/
For Docker and advanced setup options, see:
Demo
The current public demo is intended to show:
- local onboarding and coordination flow
- workbench interaction
- integration of local test bots
- a reproducible starting point for public evaluation
It is not intended to fully demonstrate all production-scale properties of Avernet, such as large-scale connection envelopes, permission isolation, audit depth, failure recovery, or long-horizon organizational collaboration.
Why Avernet
As agent systems scale, teams often hit the same four bottlenecks:
- Cannot find — capabilities are hard to discover
- Cannot align — apparent consensus hides real misalignment
- Cannot run fast — execution depends on human relay
- Cannot retain — knowledge does not accumulate as organizational capability
Avernet is built to address these problems with infrastructure for persistent agents, structured coordination, governed execution, and compounding organizational memory.
Architecture
+----------------------------+ +----------------------------+ +----------------------------+
| Local Agents | | Agent Runtime | | Existing Bot Platform |
| Plugin mode | | /ws/bot runtime | | Downlink gateway |
+-------------+--------------+ +-------------+--------------+ +-------------+--------------+
| | ^
| | |
+---------------+---------------+ |
| agent -> BCS: | BCS -> platform:
| connect / register / receive / report | dispatch / schedule / callback
v |
+----------------------------------------------------------------------------+ +-------------------+
| Avernet / BCS | | bcs-cli / tools |
| connection / registration / routing / delivery / sessions |<--->| onboard / inspect |
| collaboration state / multi-bot network management | | |
+----------------------------------------------------------------------------+ +-------------------+
Integration
Avernet does not lock you into a single agent engine. It supports two integration paths for connecting agents, runtimes, and existing bot platforms into one collaboration network.
| Integration path | Best for | Current capability | Docs |
|---|
| Plugin integration | OpenClaw, DeepSeek Harness, local agent runtimes, custom bot processes | Agents actively connect to Avernet through a plugin or runtime for registration, onboarding, message receiving, and result reporting. | Bot Integration Guide, Local OpenClaw from source, DeepSeek Harness plugin |
| Gateway integration | Existing bot platforms, multi-instance agent services, external scheduling systems | Avernet dispatches tasks to an external platform, which schedules agents and reports results back when work completes. | Bot Platform Integration |
Repository layout
Avernet/
├── apps/
│ ├── bcs/ # Bot Coordination Service (Rust) and channel plugins
│ ├── backend/ # Backend services (Python)
│ ├── gateway/ # External API gateway
│ ├── frontend/ # Web workbench
│ ├── frontend-nextgen/ # Next-generation workbench
│ ├── evolverun/ # AgentEvolve and TaskGuard
│ └── ... # baas, bcsfuse, proxy
├── engine/ # Engine adapter
├── middleware/ # Python and Rust middleware
├── singlebox/ # Local development stack (singlebox.sh)
├── docker/ # Container images and deployment scripts
├── docs/ # Guides, ADRs and architecture rules
├── AGENTS.md # Rules for contributors and AI coding agents
├── CONTEXT-MAP.md # Entry point for architecture and module context
├── README.md
└── README.zh-CN.md
Documentation
Security
Do not commit secrets, tokens, cookies, private keys, private service endpoints, local databases, runtime logs, or machine-specific configuration.
If credentials have already been committed, revoke or rotate them before cleaning repository history.
License
This project is licensed under the Apache License 2.0.
