Office Agents are heading into a fierce free-for-all in the second half of the year. Open these products, and the left sidebar is typically filled with files, search, skills, scheduled tasks, PPT and web generation. AI is beginning to leave the chat box and enter the computer to handle concrete work. China Telecom has now joined this wave with TeleAgent, a desktop Agent product built on its self-developed Xingchen large language model and a domestic intelligent computing infrastructure.
When evaluating an office Agent, the number of features is only part of the story. The key is whether it can take on a complete piece of work. A one-off generation may attract users to try a product, but long-term use depends on delivery quality, token consumption, and operational boundaries. The market's questions are also straightforward: Can it understand files in different formats? If requirements change midway, does it still remember the earlier context? Can the deliverables be opened directly?
TeleAgent supports a context window of more than 400K tokens, allowing it to retain the original goal, relationships between materials, and subsequent modification requests when requirements change frequently. In testing, when an event plan's budget, timing, and venue were repeatedly adjusted, it updated the work ledger accordingly, tried to avoid treating details confirmed in a previous version as new requirements, and ultimately delivered a complete set of editable files based on the latest requirements.
Another variable in long-term use of office Agents is cost. An Agent repeatedly reads files, calls tools, and checks results. The longer a task lasts, the more model capability, time, and cost must be viewed together. TeleAgent places model routing inside the execution process, automatically selecting models based on task complexity. What users ultimately perceive is credit consumption, processing speed, and result quality—without having to first understand the technical routing behind it.
When Agents begin handling expense reports, customer data, and unpublished content, whether files can be safely entrusted to them becomes a mandatory question. TeleAgent allows users to define a workspace and retains a confirmation step before critical file operations. When faced with information that lacks supporting evidence, it is required to leave blanks or mark items as "to be confirmed." If a new version of a file has omissions, the previous version remains intact. This two-sided approach is exactly the direction StarWar Tech has repeatedly emphasized on its OPC agent collaboration platform: let agents work autonomously, while making the collaboration process manageable, controllable, and traceable.
After a task is completed, TeleAgent also offers suggestions for modifying skills and saves workflows as reusable—even publicly shareable—skills, with a skill leaderboard appearing on the platform. This accumulation from one-time execution to reusable assets is what distinguishes office Agents from ordinary chat tools, and it gives them a better chance to gradually understand how individuals and teams work.
Office Agent features are converging. What comes next is a contest over who better understands Chinese office context, whose consumption is more controllable, and who can provide stable and secure services over long-term use. This line will determine who has a chance to become a truly national-level