Skip to content

AgentScope Runtime

★ New
assess
AI / ML open-source Apache-2.0 open-source

At a Glance

Python FastAPI-based agent deployment runtime by Alibaba's Tongyi Lab with five sandbox types, Agent-as-a-Service streaming APIs, multi-framework adapters (LangGraph, Agno, Microsoft), and nine deployment targets from local daemon to Kubernetes and Alibaba Cloud.

Type
open-source
Pricing
open-source
License
Apache-2.0
Adoption fit
small, medium
Top alternatives

What It Does

AgentScope Runtime is a Python package maintained by Tongyi Lab (Alibaba Inc.) that provides two tightly coupled functions: an agent deployment engine and a sandboxed tool execution environment. The engine wraps FastAPI (via direct class inheritance since v1.1.0) to expose agent logic as streaming HTTP APIs using Server-Sent Events, with built-in session management, conversation history, health monitoring, and a Distributed Interrupt Service for pausing and resuming agent tasks mid-execution. The sandbox module provides isolated environments for tool calls, covering five types: Base (Python/shell), GUI (desktop), Browser (web automation), Filesystem, and Mobile (Android emulation), each available in both synchronous and async variants.

The runtime is the production-focused complement to the AgentScope Python framework, also by Alibaba, and includes adapters for deploying agents built with LangGraph, Microsoft Agent Framework, and Agno on the same infrastructure. The overall design philosophy is “white-box” — the full execution context (prompts, API calls, memory, sandbox) is visible and configurable rather than abstracted away.

Key Features

  • Agent-as-a-Service (AaaS) API: @agent_app.query() decorator converts agent logic into a production FastAPI endpoint with automatic SSE streaming, health checks, and lifecycle management
  • Five sandbox types: BaseSandbox, GuiSandbox, BrowserSandbox, FilesystemSandbox, MobileSandbox — each with sync and async variants; Docker + optional gVisor for local use, Kubernetes containers for production
  • Nine deployment targets: Local Daemon, Detached Process, Kubernetes, ModelStudio, AgentRun, PAI (Platform for AI), Knative, Kruise, and Function Compute (FC) — last five are Alibaba Cloud-native
  • Distributed Interrupt Service: Runtime task preemption with developer-configurable state persistence and recovery logic; introduced v1.1.0 (February 2026)
  • Multi-framework adapters: Wraps agents built with LangGraph, Microsoft Agent Framework, and Agno (AutoGen in progress) without requiring agent code rewrites
  • A2A protocol support: A2AFastAPIDefaultAdapter for Agent-to-Agent protocol communication with a built-in service registry for agent discovery
  • OpenAI SDK compatibility mode: Drop-in API compatibility layer for existing OpenAI SDK clients
  • OTel-compatible observability: Distributed tracing and per-session logging designed for OpenTelemetry-compatible backends
  • Session persistence: Redis or in-memory session state, configurable per deployment target

Use Cases

  • Alibaba Cloud AI deployments: Teams on Alibaba Cloud wanting production deployment of agents on PAI, ACK, or Function Compute with minimal custom infrastructure
  • Multi-framework shops: Organizations running a mix of LangGraph and Agno agents who want a single deployment and sandbox runtime rather than framework-specific hosting solutions
  • Sandboxed tool execution at scale: AI agents that need to execute shell commands, manipulate files, or automate browsers in isolated containers with a consistent API across environments
  • Agents requiring runtime interruption: Workflows where human oversight requires pausing a running agent task, persisting its state, and allowing re-entry after a decision — the Distributed Interrupt Service addresses this directly

Adoption Level Analysis

Small teams (<20 engineers): Fits for teams already on Alibaba Cloud or building on the AgentScope framework directly. The pip install agentscope-runtime entry point and decorator-based API are genuinely low-friction. Caution: the API broke between v1.0 and v1.1.0 (factory pattern deprecated), and the project launched in December 2025, meaning there is very limited community knowledge, StackOverflow coverage, or battle-tested examples outside Alibaba’s own documentation.

Medium orgs (20–200 engineers): Fits with significant caveats. Framework-agnostic adapters for LangGraph and Agno are a differentiating capability for organizations already invested in those frameworks. However, all Docker images are hosted on Alibaba Cloud Container Registry (not Docker Hub), creating a supply chain dependency. Deep deployment features (PAI, AgentRun, Kruise) are Alibaba Cloud-specific and provide no value on other clouds. Consider this primarily if your cloud strategy already includes Alibaba Cloud.

Enterprise (200+ engineers): Limited fit outside Alibaba Cloud ecosystems. The framework is too young (< 6 months at GA) for large organizations requiring API stability commitments, SLA-backed support, or multi-year roadmap visibility. Enterprises outside Alibaba Cloud are better served by LangGraph Platform, Agno’s AgentOS, or a custom deployment on Kubernetes using E2B or OpenSandbox for tool isolation.

Alternatives

AlternativeKey DifferencePrefer when…
Agno (AgentOS)Native stateless FastAPI runtime for Agno agents; richer HITL and approval workflowsYou are building new agents and want batteries-included deployment without Alibaba Cloud dependency
LangGraph PlatformFirst-party hosted or self-hosted deployment for LangGraph agents with durable executionYour agents are LangGraph-native and you want the deepest integration with checkpointing and human-in-the-loop
OpenSandboxSelf-hosted Alibaba-origin sandbox with multi-language SDKs, focused on code execution isolationYou want only the sandbox component without the deployment runtime
DaytonaLightweight open-source Docker-based sandbox, sub-90ms creation, Computer Use supportYou need fast ephemeral sandbox creation without a full agent deployment framework
E2BManaged Firecracker microVM sandbox, sub-200ms cold starts, fully hostedYou want managed sandboxes with no infrastructure to operate

Evidence & Sources

Notes & Caveats

  • API instability at v1.x: The v1.0 to v1.1.0 transition deprecated the factory pattern in favor of direct FastAPI inheritance — a non-trivial migration for any existing v1.0 adopters. The project is stabilizing its API surface but is not yet at the stability level expected for a “production-ready” framework.
  • Alibaba Cloud Registry dependency: All Docker sandbox images are pulled from Alibaba Cloud Container Registry (registry-intl.aliyuncs.com). Teams in regions with connectivity restrictions to Alibaba infrastructure, or with supply chain security policies requiring Docker Hub or self-hosted registries, will need to re-tag images manually.
  • Alibaba Cloud coupling: Five of the nine deployment options are Alibaba Cloud-specific (ModelStudio, AgentRun, PAI, Kruise, Function Compute). The remaining four (Local, Detached Process, Kubernetes, Knative) are cloud-agnostic, but the richest operational features are tied to Alibaba’s platform. Evaluate honestly whether this is “broad deployment support” or “Alibaba Cloud deployment with K8s/Knative as fallback.”
  • No independent production evidence: As of April 2026, no publicly documented production deployments from teams outside Alibaba exist. The HiClaw/CoPaw case study is the only documented external adopter, and it is an Alibaba-sponsored ecosystem partnership rather than an arm’s-length evaluation.
  • Parent framework maturity: The AgentScope parent framework (separate package) has a longer history and a peer-reviewed paper. AgentScope Runtime is a younger, separate package focused on deployment. Teams evaluating AgentScope Runtime should assess both components together.
  • No security audit: The “hardened sandbox” claim for tool execution has not been verified by an independent security audit. The Docker + gVisor approach is industry-standard, but the sandbox server code itself has not been publicly reviewed for escape vectors.
  • Celery mode limitation: In Celery-based deployment mode, only the final response is stored; intermediate streaming events are discarded. This is a documented regression for streaming-dependent agent workflows.

Related