AgentScope 1.0: A Developer-Centric Framework for Building Agentic Applications

AgentScope is Alibaba's open-source framework for building agent applications, and this paper describes its 1.0 redesign. It is organized in three layers.

The first layer is four foundational components. A message is a sender name, a role and a list of typed content blocks (text, image, audio, video, tool use, tool result, thinking), with an ID and timestamp. The model layer hides provider differences behind formatters and a single response schema, tracks tokens and latency per call, and emits OpenTelemetry traces. Memory has a short-term buffer and a long-term interface, with an implementation built on Mem0. The toolkit turns Python functions and MCP servers into JSON schemas the model can call.

The second layer is the agent. AgentScope recommends one architecture, ReAct, and gives every agent three entry points: reply to reason, act and answer; observe to take in information without answering; and handle_interrupt to react when a user stops it. Around that loop it adds parallel tool calls, tools the agent can switch on and off in groups, hooks before and after every step, and state that can be saved and restored.

The third layer is developer tooling: an evaluation module with a sequential evaluator for debugging and a Ray-based one for scale, a studio that renders traces and evaluation results, and a runtime that deploys an agent as a FastAPI service and runs tools in sandboxes.

The paper ends with built-in agents for deep research, browser use and multi-step planning. It contains no experiments.

When a customer pilot moves from a notebook to people using it, the questions change. Users interrupt: "stop, check the other account first." Agents grow from five tools to forty. A browser tool needs to stay logged in between calls. A security team asks where tools run. The product manager asks whether last week's prompt change helped.

AgentScope is useful because it names these needs as framework features, so you can use it as a checklist when you choose or build a runtime.

Interruption keeps partial work. When a user cuts in, whatever the model or tool already produced is kept in memory with a note saying it was interrupted. The next reasoning step can resume, revise or change course instead of starting blind.

Tools come in groups. Only active groups go into the prompt, and the agent switches groups itself. Fewer tool descriptions per call means fewer tokens and fewer wrong picks.

Sessions are a choice. Stateless MCP clients reconnect for every call; stateful ones keep a session, which a logged-in browser needs.

Evaluation reports a distribution. The studio shows repeated-run results with bootstrapped confidence intervals and sorts test items into consistently right, consistently wrong and unstable.

Because the paper reports no measurements, treat each feature as a claim about design, not about results. The parts worth copying are the interfaces; whether they hold up under your load is for you to measure.