China Telecom's TeleAgent lands in IDC's top three for enterprise general-purpose agents

China Telecom's TeleAgent has landed in the top three of China's first technical evaluation of enterprise general-purpose agents, published by IDC. The evaluation, reported by QbitAI on September 17, deliberately avoided the usual large-model benchmarks and instead used nearly 100 undisclosed tasks to test whether an agent can actually finish real office work end to end.

The scores: 3.49 points on routine tasks and 3.36 on complex tasks. Across nine capability dimensions, TeleAgent took a full 5 points for task performance and the highest marks in the evaluation for cost efficiency.

The product is young. TeleAgent's desktop version was still in internal trial in April this year, and V1.0 only launched publicly in July. In a little over a month, its user base has already approached 1.2 million.

IDC defines an enterprise general-purpose agent as something that can understand a goal, invoke tools and skills, connect enterprise knowledge and business systems, and keep executing. Its roughly 100 undisclosed tasks mix everyday work — email, scheduling, document handling — with long-context, multi-step loops, complex slide decks, spreadsheet processing, and browser operations. Some are close to real work: cleaning a dirty dataset of 3,755 rows, or distilling roughly 30,000 characters of material into an 11-page deck. Finishing was not enough; IDC also verified system state and final files to confirm the tasks were genuinely completed.

The results show most general-purpose agents can now run complex tasks, but as tasks lengthen and step counts rise, the gap in delivery quality and resource consumption widens. IDC singles out a term for this: the agent harness, the execution system built around the model. The model understands and reasons; the surrounding system budgets context, schedules tools, persists task state, controls the execution environment, decides whether to retry, reroute, or resume after an exception, and checks whether the job was actually done. Once tasks get long, that is where the differences show up.

TeleAgent's high scores can be traced along that engineering chain. On context, it lightly prunes low-value older tool output, and if the window is still too long, a dedicated compression model takes over; it also monitors how close it is to the limit and compresses before continuing, with a context window now past 400K. Its autoDream long-term memory periodically consolidates recent conversations and merges them with existing memory, so project background, habits, and preferences survive across sessions.

On cost, TeleAgent built its own ModelRouter: each request is classified by modality, length, and keywords, then routed to a lightweight, balanced, or flagship tier. Translation, summarization, and formatting go to the cheap tier; slide and document generation and office automation run on the balanced tier; deep research, code debugging, and complex planning reach for the flagship. By the reported numbers, that routing cuts inference cost by roughly 40%, keeps simple-task response times at 3-5 seconds, and adds under 10ms of routing latency, alongside PD disaggregation, KV cache, and load-aware scheduling.

Security explains why the launch was slower than an internet product's would be. Skills are checked for provenance, viruses, and vulnerabilities before entering the system; short- and long-term memory are managed separately; file access is confined to a designated working directory; high-risk commands are monitored; and code and file operations run in isolated environments wherever possible. China Telecom spent two extra months on security testing and capability tuning before the July launch. QbitAI reports TeleAgent is among the first agent products to pass the relevant China Academy of Information and Communications Technology security evaluation, with a 98.2% pass rate in China Telecom Research Institute's internal security testing and 99.7% across 28 scenarios and 336 test cases with external security vendors.

The groundwork predates this year's agent boom: China Telecom began building the Xingchen agent platform at the end of 2024, designated it group-wide agent infrastructure in April 2025, and shipped a web version of Xingchen Super Agent that August, later layering in coding enhancements, skills, and a desktop environment. Its internal skill marketplace now lists 56,000 skills with nearly 2 million additions, contributed by 20,000 employees. IDC's April survey found 48.5% of enterprises already evaluating, piloting, or using general-purpose agents, and 96.9% of enterprise agent applications touching office automation. As underlying models stop explaining the whole gap, the advantage increasingly sits in the engineering system around them.