China Telecom Launches TeleAgent Desktop Agent with 400K Token Context

In the second half of 2026, competition among office agent products intensified markedly. China Telecom launched TeleAgent, a desktop agent product based on its proprietary Xingchen large model and a domestic AI computing foundation, supporting context windows exceeding 400K tokens. TeleAgent can read files in various formats and invoke tools to complete tasks such as data processing, PPT generation, and webpage creation. Users can restrict the working space, retain a confirmation step before critical file operations, and previous versions of files are fully preserved.

In an event-planning test, TeleAgent used project materials to generate an activity plan for the simulated brand "Qike Coffee," producing a project Excel file, a one-page execution brief, and an editable 8-page PPT, while also establishing a work ledger. During the test, requirements were adjusted multiple times: the event name was changed to "Qike Friday After-Work Party," attendance was reduced from 160 to 100, and the budget was cut from RMB 120,000 to RMB 80,000, prompting TeleAgent to trim giveaways and lower refreshment costs. The start time was postponed from 14:00 to 16:00, with corresponding adjustments to check-in, rehearsal, staff arrival, and material setup. Venue and PPT color schemes were also revised per the latest requirements. Item-by-item verification confirmed that the final deliverables matched every round of modifications.

In a student-schedule test, TeleAgent processed syllabi for six courses, class schedule screenshots, course notices, club duty rosters, and internship shift information. It output a semester task list, a 16-week plan, and a browser-openable "This Week's Task Board," with each task annotated by date, grading requirements, and source. For a final presentation time not yet confirmed, the product did not fabricate a guessed date. The user research course presentation was listed as October 29 in the syllabus, while a group notice proposed moving it to the 30th pending academic affairs confirmation; TeleAgent retained both dates and noted conflicts with the project management course and internship, respectively.

In a product-launch preparation test, the simulated team of five aimed to recruit the first batch of beta users for "Shixu." The folder contained founder notes, 10 user interviews, a feature list, a cost table, a logo, and two sketches. TeleAgent first reconciled discrepancies in old information, including early notes citing 100 recruits versus an officially approved quota of 50, as well as changes to pricing and launch date. It then delivered, based on confirmed information, a product positioning statement, a recruitment plan, a two-week action table, a 6-page PPT, and a webpage prototype. The action table specified responsible persons, completion dates, and task arrangements; interview insights were tagged with interview numbers; features not in the initial release scope were not presented as available. The webpage could be opened locally, with forms alerting users to missing names and incorrect email formats, and it was marked as a demo page.

The test also simulated last-minute changes: the beta quota was cut from 50 to 30, pricing was set to "not yet disclosed," and a note stating "product data is stored locally only" was added. TeleAgent preserved the first version of files while generating a second set, updating the PPT and webpage to reflect 30 slots, marking the original price position as "pricing to be announced," and adding the local data storage note to the product description. TeleAgent consolidated scattered materials into an executable launch package. The product uses context engineering to retain initial goals, material relationships, and subsequent modification requirements across long tasks, and employs a Model Router to automatically select models based on task complexity. Registered users receive 3,000 credits immediately, with additional credits available daily. OpenAI's Codex usage study, published in June 2026, indicates that agents are shifting the basic unit of knowledge work from short conversations to delegated tasks lasting tens of minutes or even hours.