Alibaba Qwen Office Open-Sources MyContext to Give Agents Work Context

Although Agents are improving in task orchestration and tool invocation, they often fail to understand workflows in real business scenarios because key information is scattered across emails, IM, documents, and databases, with issues like real-time updates, version conflicts, and permission boundaries.

A Confluent 2026 survey shows that 66% of enterprises believe data infrastructure and data quality are slowing down Agentic AI adoption, and 80% have listed 'using their own data to drive AI' as a business priority. This indicates that the bottleneck for enterprise AI adoption has shifted from model capabilities to data infrastructure.

Alibaba Qianwen Office's open-sourced MyContext is positioned as context infrastructure, aiming to process scattered, heterogeneous personal work data into dedicated profiles that Agents can understand. It supports IM, documents, meetings, business collaboration records, and local and other work data sources, continuously aggregating and organizing them with user authorization to form dynamically updated work profiles.

MyContext emphasizes traceability, with each conclusion retaining an evidence chain that can be traced back to original chats, documents, or meeting records; it is also constrained by user and organizational permissions. This avoids 'black-box memory' and allows Agents to understand real business context.

Temporal data handling: MyContext does not process data simply in chronological order. Instead, it binds each piece of raw information to a stable source identifier, with idempotency based on the data source identifier. Even if timestamps are old, as long as they haven't been consumed, they are still processed. It also uses conversation idle intervals as Session boundaries, making context segmentation align with real interaction rhythms.

Fact conflict handling: A three-state merge mechanism is used. Consistent information boosts confidence, supplementary information is merged into existing conclusions, and genuine conflicts preserve multiple facts while lowering confidence, explicitly exposing them to users. Human-confirmed conclusions are given higher priority, and models are prohibited from automatically overriding them.

Cost control: Incremental computation is used, where rules that can be determined locally are processed first, and only ambiguous information is sent to the model. Already-computed results are reused, and multiple updates are batched for processing, combined with version caching, batch triggering, and tiered degradation strategies to reduce redundant computation.

MyContext is not only for individuals but also extends within organizations, forming a closed loop of 'data aggregation - context processing - Agent consumption' with DingTalk and Qianwen Office. DingTalk has over 20 million enterprise organizations and nearly 800 million users, making it a natural data entry point. The Qianwen Office team's expertise in heterogeneous data processing and context engineering can help enterprises integrate knowledge from systems like DingTalk, Feishu, Salesforce, and SAP.

This closed loop aims to let enterprise scattered data settle into trusted, traceable, and continuously evolving context, transforming Agents from tools that passively receive instructions into long-term productivity units that understand the organization, continue tasks, and participate in collaboration.