Highlights and PPTs from the Guangzhou Agent Open‑Source Developer Salon

Agenda 1: From HiClaw to AgentTeams – Multi‑runtime Worker and Enterprise Collaboration Upgrade

AgentTeams is positioned as a foundation for multi‑agent collaboration and unified management, addressing coordination challenges in complex task delivery. It organizes humans and heterogeneous runtimes (e.g., QwenPaw, OpenClaw) through unified Team/Room/Task semantics, enabling division of labor, collaboration, intervention, and acceptance. Its core relies on a Matrix‑based collaboration architecture and the TeamHarness unified protocol, ensuring observability and intervene‑ability at any time. AgentTeams complements ecosystem components such as AgentScope to form a complete development‑runtime‑governance chain.

Agenda 2: Building Distributed Enterprise‑Grade Agents with AgentScope 2.0 Harness

AgentScope 2.0 is an enterprise‑oriented, distributed agent development framework. Its core is the Harness design built on ReActAgent, encapsulating engineering capabilities required for long‑running agents. Features include a Workspace that files agent state, a two‑layer long‑term memory that automatically aggregates cross‑session facts, context compression to stay within token budgets, a planning mode that follows “think‑plan‑confirm‑execute”, and sandbox management that isolates dangerous operations. Different configurations let Harness quickly construct personal assistants, multi‑tenant platforms, data agents, and other enterprise‑level applications.

Agenda 3: Nacos Adds Agentic Resource Discovery (ARD)

Nacos 3.3 introduces Agentic Resource Discovery (ARD), a unified resource search engine designed for AI applications. ARD breaks the silos between Skills, MCP, and Agents by combining keyword and vector retrieval to enable cross‑type resource discovery. Its architecture stores resource facts first, then builds indexes asynchronously to guarantee eventual consistency. The release also enhances Remote Agent Discovery (RAD) and single‑resource authorization, moving toward a unified AI resource management hub.

Agenda 4: OBI – Non‑intrusive AI Agent Observability

OBI (OpenTelemetry eBPF Instrumentation) provides a non‑intrusive observability solution for AI agents. Using eBPF, it intercepts network traffic without modifying application code and generates standard traces and metrics that include GenAI semantics such as model, tokens, and tool calls. OBI covers raw HTTP calls, cross‑language coroutines, and sandbox container calls, addressing gaps in traditional APM for AI workloads. Real‑world cases demonstrate its ability to locate poor RAG retrieval quality and abnormal token consumption.

Agenda 5: UnifiedModel – A Semantic Layer for AI Agents

UnifiedModel proposes a “semantic layer” to solve the “blind‑fold” problem of agents lacking understanding of objects and relationships in real systems. It defines unified primitives—Set (object), Link (relationship), Field (attribute)—to organize scattered data into an object graph that agents can comprehend. The layer offers a unified SPL query language, abstracts storage differences, supports progressive disclosure and MCP protocol, and enables agents to dynamically discover and invoke capabilities. Experiments show that adding the semantic layer markedly improves accuracy on data‑question answering tasks, making it essential for intelligent reading and root‑cause analysis scenarios.

Agenda 6: Vibe Chaos – Production‑grade Agent Engineering Practices

Blade AI advances chaos engineering from “Vibe Coding” to “Vibe Chaos”, emphasizing that production agents must have engineered safety nets. Four mechanisms are introduced: code‑level guardrails that enforce deterministic checkpoints, frozen contracts that lock experiment goals after user confirmation, independent recovery state machines that reliably clear failures, and independent verification that validates fault impact via an evidence chain rather than trusting the executor’s claim.

The event also featured a hands‑on session where speakers deployed AgentTeams, walked participants through simple scenarios, and engaged in interactive Q&A.