How DeepScience and Alibaba Cloud’s AgentRun Accelerate AI Research Agents at Full Speed

AI and intelligent agents are reshaping the fundamentals of scientific research, requiring infrastructure that embraces an AI‑native paradigm. Alibaba Cloud’s AgentRun combines Serverless’s limitless elasticity, zero‑ops management, and pay‑per‑use pricing with deep integration into AI‑native workloads, allowing researchers to focus on problem definition and innovative decision‑making.

Requirements and Challenges for Scientific Agents

According to DeepScience R&D lead Lu Jintan, scientific work demands absolute precision—eliminating model hallucinations and ensuring data source accuracy—while spanning the full lifecycle of reading literature, performing simulations, and automating experiments. This necessitates agents capable of handling long‑running asynchronous tasks that may last hours or days. Consequently, scientific agents impose higher demands on infrastructure flexibility, security isolation, and end‑to‑end observability.

AgentRun’s Extreme Elasticity

AgentRun is built on Function Compute (FC) and supports elasticity from zero to millions of concurrent executions. Its "shallow sleep" technique preserves rapid response while delivering performance gains hundreds of times higher than traditional servers, and its "deep sleep" persists session state for extended periods, charging only when agents actively execute tasks. These features collectively lower total cost of ownership by about 60%.

Breaking Serverless Stateless Limits

To address the long‑duration, interactive nature of scientific simulations, AgentRun introduces a session‑affinity mechanism that routes consecutive requests for the same research topic to the same instance, providing a persistent, stateful context for multi‑turn dialogues. Snapshot technology further enables agents to "deep‑sleep" with their state persisted, allowing second‑scale wake‑up when results are ready.

Secure Sandbox Execution

AgentRun Sandbox mitigates the security risks of code execution by leveraging MicroVM kernel‑level isolation, multi‑dimensional compute isolation, and dynamic mounting. Built‑in tools such as Code Interpreter and Browser Tool offer out‑of‑the‑box multi‑language execution environments, while fine‑grained permission controls and audit logs ensure that security does not hinder large‑scale agent deployment.

Full‑Chain Tracing for Debugging and Optimization

Given the lengthy decision chains of scientific agents, pinpointing failures is difficult. AgentRun’s full‑chain tracing visualizes the entire request‑to‑response flow, reconstructing each step—intent understanding, model inference, tool invocation, and knowledge retrieval—along with status and latency, thereby making the decision process transparent and significantly reducing debugging effort.

Ecosystem Expansion and Rapid Tool Onboarding

DeepScience launched the MCP market on its Bohrium research space, onboarding over 50,000 scientific tools within two days using minimal configuration. AgentRun’s open ecosystem supports rapid integration of frameworks such as AgentScope, LangChain, ADK, and CrewAI, covering no‑code, low‑code, and full‑code development scenarios. The Serverless nature enables swift creation of MCP servers, and the AI Gateway converts traditional HTTP APIs into AI‑Agent‑compatible endpoints.

Future Outlook

DeepScience and Alibaba Cloud plan to deepen collaboration, aiming to provide every researcher with a secure, efficient, and universally accessible cloud foundation, allowing scientists to concentrate on discovery while AI agents become the core drivers of cross‑industry solutions.