Inside DeepModeling's Research Agent Observability Stack with AgentLoop

Research Agent Infrastructure Challenges

Research tasks require literature and data retrieval, professional tool invocation, scientific computation, and result analysis and verification. Unlike general-purpose agents, research agents must continuously manage complex cross-tool, cross-environment execution processes and retain sufficient evidence to support conclusions.

DeepModeling's AI for Science Ecosystem and Agent Infra

DeepModeling builds research agent products and infrastructure around AI for Science . Its ecosystem includes SciMaster for general research tasks, MatMaster and BioMaster for materials and biology domains, and delivery surfaces such as Bohr Scientific Navigation, Bohr Scientific Space, and Bohr-CLI that integrate literature research, data processing, scientific computing, and result delivery into researchers' daily workflows. A candidate material screening workflow demonstrates the typical process: retrieve literature and data, design and run computations, analyze and verify results, deliver candidates with supporting evidence, and continue verification when gaps appear. MatMaster's public demo shows candidate structures, analysis reports, calculation inputs, and result files as delivery artifacts.

Although domain methods differ, tool invocation, runtime state management, long-task execution, and evidence retention are common infrastructure challenges. DeepModeling abstracts these shared capabilities into Agent Infra . Tasks enter from research applications, web, or API; Shrimp Runtime orchestrates model judgment, tool execution, context management, and continuous execution; Bohr-CLI provides a unified command entry for research capabilities, while the runtime can adopt built-in tools or MCP as needed.

Three Core Challenges for Research Agents

Black-Box Execution

Research agents may invoke literature parsing, knowledge retrieval, modeling simulation, or experimental data analysis. An anomaly in any tool, parameter, dependency environment, or returned result can affect subsequent judgments. Without unified runtime observation, it is difficult to quickly identify whether a problem originates in model judgment, tool invocation, or the external compute environment.

Long-Running Scientific Computation Requires Persistent State Management

Scientific computing includes both online code execution needing immediate feedback and offline jobs submitted to run independently and queried by task ID. Tasks may experience waiting, service interruption, page disconnection, recovery, or cancellation. For a research agent, a single tool invocation ending does not mean the overall task is complete; external job status, execution steps, and artifact locations must be continuously verifiable.

Execution Success Does Not Equal Scientific Validity

Research tasks demand far higher verification standards than ordinary information generation. A command running successfully only proves minimal executability; whether the computation satisfies numerical error bounds and constraints, and whether the final deliverable meets the research goal, requires further judgment combining parameters, logs, raw results, scoring items, and failure feedback.

Agent Infra Implementation with AgentLoop Integration

Architecture and Division of Labor

In the architecture, Alibaba Cloud AgentLoop handles runtime observation of key invocations, latency, and traces. Durable execution, asynchronous jobs, workspace artifact retention, and scientific evaluation are handled by existing Infra mechanisms. Teams can first identify which step anomalies concentrate on via the call chain, then combine business task, external job, and artifact records to confirm actual state, avoiding judgment based solely on a single invocation result.

This division is critical for research scenarios: runtime observation answers "what happened, how long, where anomalies concentrate"; business records answer "what state are the task and external jobs in now"; scientific validation answers "do the computation and deliverables meet research goals". The three must connect but cannot replace each other.

Capability Supply Side

On the supply side, "read, compute, do" tools — literature parsing, knowledge retrieval, scientific computing, experimental data analysis — complement open-ecosystem research software. Online sandboxes support interactive execution and immediate feedback; offline compute tasks suit long-running, batch execution; cloud workspaces persist inputs and research artifacts so files are not lost when temporary execution environments are released.

Integrating AgentLoop for Runtime Observability

A tool invocation ending does not mean the research task is complete. Facing "task not completing for a long time", the team needs to distinguish tool invocation errors, external compute still waiting, or frontend not continuing updates — three phenomena with different verification paths. This motivated AgentLoop integration.

Three Troubleshooting Paths

Tool invocation error : further verify failed invocation, input parameters, and dependency environment.

External compute long wait : query job status, logs, and progress by task ID.

Frontend not updating : simultaneously check server-side task status and received event cursor to distinguish whether the task is still executing or the process stream needs recovery.

Benefits of AgentLoop Integration

AgentLoop provides runtime observation for key call chains, enabling engineers to see which stage a task is running in, where anomalies appear, and combine invocation information with actual task status for analysis. For tool errors, external compute waits, or frontend stalls, teams can apply corresponding verification paths, reducing blind troubleshooting.

AgentLoop's observation collaborates with DeepModeling's existing durable execution, async task management, and workspace artifact retention to support cross-tool, cross-environment long-running workflows. Candidate structures, analysis reports, calculation inputs, and result files return to the workspace, forming a complete, analyzable, and deliverable work process.

Invocation chains, execution steps, and research artifacts together form the basis for problem analysis. Teams can locate gaps at the data, tool, agent orchestration, and model levels, then verify improvements through fixed-condition version comparisons and regression tests, shifting research agent optimization from scattered experience toward inspectable, reproducible engineering processes.

Long-Task Recovery and Cleanup Mechanisms

Research tasks often wait for external compute returns. DeepModeling uses durable execution to record completed steps, reuse existing records after service interruption, and verify external job status before continuing execution. After page disconnection, the event cursor resumes reading process information. For cancellation, the system stops associated execution, confirms resource and billing cleanup, then enters the final state.

This mechanism links task status, business steps, external task IDs, and artifact locations, enabling progress verification for long tasks and reducing risks of duplicate execution and state misjudgment.

Bohr Tool Base and Product Delivery

Core Capabilities: Read, Compute, Do

Beyond hardening Agent Infra, DeepModeling enhances the "Bohr Tool Base" to output research capabilities, summarized as "read, compute, do", plus the ecosystem tool Deploy-Master.

Read : processes papers, formulas, and knowledge clues.

Compute : performs modeling and simulation around structures and parameters.

Do : currently focuses on experimental data analysis, e.g., cryo-EM particle recognition and spectrum analysis.

These core capabilities complement open-ecosystem research tools.

Deploy-Master: Ecosystem Tool Discovery and Verification

Deploy-Master

discovers research software from the open ecosystem, builds runtime environments, wraps entry points, and executes minimal verification commands. As of January 2026, in 52,550 build attempts, 50,112 tools passed minimal execution verification — a 95.36% pass rate. This figure reflects minimal executability under the experimental conditions at that time; it does not imply numerical correctness, nor that the research task is complete.

Bohr-CLI: Unified Command Entry

Once research capabilities enter the platform, they need a clear, consistent invocation method. Bohr-CLI lets developers, scripts, and agents know what to input, what is returned, and what actions can follow. Structured results and request IDs facilitate tracing specific requests. The unified command entry is not the only channel; the runtime can still use built-in tools or MCP per task.

Product Forms and MatMaster Demo

These public capabilities reach research work through different product forms: Bohr-CLI for developers, scripts, and agents; Bohr Scientific Navigation for browser-based retrieval, paper reading, and research; Bohr Scientific Space on desktop for organizing projects, analyzing computations, and managing result files. As of architecture data verified on September 15, 2026, SciMaster local mode is still cloud-orchestrated, with tool and file operations executing in user-authorized directories.

For example, MatMaster 's public demo shows candidate structures, analysis reports, calculation inputs, and result files as delivery artifacts. These artifacts return to the workspace, forming a process that can be further analyzed and delivered. The public animation's ~9.3-second playback duration is an editing effect, not actual compute time; the demo illustrates the delivery method, while whether the scientific goal is achieved still requires further validation.

Conclusion

For DeepModeling, Agent Infra is not a single tool or model, but a common foundation supporting research agents' continuous execution, runtime observation, result validation, and product delivery. By introducing Alibaba Cloud AgentLoop, DeepModeling further strengthens runtime observation and troubleshooting capabilities, providing more solid engineering support for research agents across domains to reuse infrastructure and deliver professional results.