Avernet Multi-Agent Coordination
Infrastructure for connecting, coordinating, and operating persistent agents across heterogeneous runtimes.
Per-dimension scores and reasoning
The evidence shows read-only CI permissions, tool-free example player skills, and targeted tests for identity, authorization failures, roster integrity, and secret non-disclosure. The architecture and two integration paths provide a useful high-level account of data movement. Deductions apply because permissions are still Partial, security is Planned, and audit support remains in progress; no platform-wide confirmation policy, complete sensitive-data lifecycle, dependency vulnerability scanning, or system-level rollback procedure is shown. Dispatch and session-completion effects are reasonably explicit, while recovery evidence is limited mainly to preserved game state and an asserted recoverable Pack mechanism. Apache-2.0 establishes licensing provenance, but the supplied files do not firmly establish publisher identity or a formal attribution chain.
The README's capability status and demo boundaries are broadly consistent with the state-machine contracts in the tests. CI defines Rust architecture checks and Node build, check, and test stages, while installation and dependency documentation are indexed. Deductions apply because detailed dependency manifests, a support matrix, and actual CI results are absent from the supplied material, so availability across the whole repository cannot be confirmed statically. Failure reporting is unusually strong: authentication, forbidden access, unreadable permissions, roster changes, invalid success responses, missing state, and connection failures are distinguished, with private diagnostics suppressed.
The project identifies organizational, multi-agent, heterogeneous-runtime, and human-agent scenarios and offers both plugin and gateway integration paths. Capability boundaries are exceptionally explicit through Available, Partial, In progress, and Planned labels, plus a direct statement of what the demo does not establish. Deductions apply because the strongest trigger and phase-gating evidence concerns one undercover-game profile rather than the platform as a whole. Local scripts, Docker references, external-platform integration, and a constrained Python version support ordinary environment fit, but broader operating-system, topology, and compatibility guidance is not present here.
The README is well organized around quick start, architecture, integration, repository layout, documentation, and security, with English and Chinese entry points. It supplies basic installation steps, a demo, and links to focused guides. Known limitations and public-coverage status are documented thoroughly, and the complete Apache-2.0 license matches package metadata. Deductions apply because Avernet, the package name ocb, and the component name BCS coexist without a documented naming transition; no visible FAQ, release policy, or changelog is supplied. Maintainers, support channels, contribution ownership, and upgrade responsibility are not clearly identified, and publisher identity remains unverified.
A workbench, five local test bots, structured coordination flows, and machine-readable errors make the available output useful for ordinary evaluation and integration. Combining discovery, routing, team formation, orchestration, and evaluation offers plausible value beyond isolated-agent demos. Deductions apply because the production impact and 90%+ completion claim are README assertions without supporting measurements, and several central capabilities are not yet public. Resource usage, operating burden, latency, scale limits, and monetary cost are not quantified, leaving cost-benefit only thinly supported.
Capability labels, architecture, installation entry points, licensing, CI definitions, and several security and state-machine behaviors trace to concrete supplied files. Tests and workflows partially corroborate the README's implementation claims. Deductions apply because the internal-production, 12-business-group, and 90%+ completion claims have no attached report, dataset, or independent source, while linked documentation contents were not supplied. Fact/inference separation is strong: public, partial, planned, and internal capabilities are distinguished, and the demo's evidentiary limits are explicitly stated.
- Do not treat the Ant Group production use, 12-business-group coverage, or 90%+ completion rate as independently verified; the supplied material provides no metric definition, sample size, or report.
- Public permissions are only Partial, while security and governance are Planned and audit and lifecycle support remain in progress. Perform independent threat modeling and permission review before handling sensitive or privileged workloads.
- Installation involves system packages, Rust tools, and npm dependencies, but no vulnerability-scanning or supply-chain policy is shown, and the Node CI explicitly uses --no-audit.
- Platform-wide rollback, disaster recovery, compatibility, and upgrade ownership are not documented here; validate recovery objectives and the maintenance path before production adoption.
What does this agent do, and when should you use it?
Avernet is a self-hostable, open-source infrastructure layer for persistent, heterogeneous, coordinated multi-agent systems. Its Avernet/BCS core manages connections, registration, routing, delivery, sessions, collaboration state, and multi-bot networks, while a frontend workbench provides the human interaction surface. Agents can connect actively through a plugin or the /ws/bot runtime, or Avernet can dispatch work through a downlink gateway to an existing bot platform and receive callbacks when execution completes. The repository also describes AgentEvolve for diagnosis, repeatable Bench evaluation, optimization, and recoverable Pack versions, plus TaskGuard for workflow orchestration. The recommended local stack launches the Avernet process, frontend workbench, and five test bots; the public demo is explicitly narrower than the system's claimed production deployment.
The flow begins when a bot or agent joins through plugin mode or the /ws/bot runtime and performs connect, register, receive, and report operations against Avernet/BCS. BCS then maintains connections and registrations while handling discovery, relationships, team formation, routing, delivery, sessions, and collaboration state. For an existing bot platform, BCS uses its downlink gateway to dispatch and schedule work; the external platform runs the selected agents and returns a callback and result. bcs-cli/tools provides onboard and inspect operations, and the frontend workbench supplies an interactive interface. AgentEvolve diagnoses bots, runs repeatable Bench evaluations, performs goal- or diagnosis-driven optimization, and creates recoverable Pack versions, while TaskGuard supplies workflow orchestration.
- An engineering team that wants a self-hosted environment for running several local bots and testing registration, discovery, routing, and collaboration from one workbench.
- A team operating OpenClaw, local agent runtimes, or custom bot processes that needs to connect them to a shared coordination network through plugins.
- A platform group with an existing multi-instance bot service or external scheduler that wants central task dispatch without replacing its own scheduling layer.
- An evaluation team that needs bot diagnosis, repeatable Bench runs, goal-driven optimization, and recoverable Pack versions.
- An organization combining human participation, heterogeneous runtimes, and long-lived multi-agent collaboration within one infrastructure boundary.
What are this agent's strengths and limitations?
- Two integration modes cover both actively connected runtimes and existing bot platforms that retain their own scheduling systems.
- The coordination layer includes discovery, relationships, team formation, routing, and collaboration rather than only one-shot task execution.
- The recommended local scripts start a backend, workbench, and five test bots, providing a reproducible evaluation setup.
- AgentEvolve combines diagnosis, repeatable evaluation, optimization, and recoverable versions, while TaskGuard adds workflow orchestration.
- The project reports production use at Ant Group across 12 business groups as of early July 2026, with over 90% task completion in measured multi-agent workflows.
- Public coverage trails the internally deployed system: permissions are partial, audit and lifecycle are in progress, and security and governance are planned.
- Context and memory are still planned, so the public repository does not yet provide the full organizational-memory vision described by the project.
- Container support is partial and cluster support is planned, creating additional validation work for larger production deployments.
- The public demo explicitly does not demonstrate large-scale connection envelopes, deep permission isolation, audit depth, failure recovery, or long-horizon collaboration.
- The supplied material gives no runtime-version matrix, authentication setup, capacity benchmarks, or concrete system-level recovery guarantees.
How do you install or deploy this agent?
The documented flow requires shell access, Git, network access, and a local filesystem. The supplied material does not specify an operating system, language runtime version, credentials, or which tools must already exist before the installer runs. Execute:
git clone https://github.com/inclusionAI/Avernet.git
cd Avernet
./scripts/singlebox.sh install-tools
./scripts/singlebox.shA Docker path is also documented by the repository, but the supplied material does not include copyable Docker commands or establish Docker as a mandatory dependency of the recommended script.
How do you use this agent?
After startup, open http://127.0.0.1:8000/ in a browser. The local stack includes the Avernet process, frontend workbench, and five test bots for exercising onboarding, workbench interaction, and coordination. For a real integration, use the plugin path when OpenClaw, a local runtime, or a custom bot process should register, receive messages, and report results directly. Use the gateway path when Avernet should dispatch tasks to an existing bot platform, which schedules execution and reports completion. The supplied material contains no concrete API request, environment-variable, or authentication example, so production integrations require additional configuration from the referenced integration documentation.
How does this agent compare with similar options?
Compared with isolated single-agent demos, Avernet targets discovery, team formation, routing, and sustained collaboration among multiple persistent agents. Compared with infrastructure tied to one agent engine, it exposes plugin/runtime and platform-gateway integration paths for heterogeneous systems. Its public demo and partially released components, however, do not establish every production-scale property claimed for internal deployments.