央企做了个通用Agent，直接杀进IDC实测前三！
Ranking
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - China Telecom’s enterprise-focused TeleAgent ranked third in IDC China’s evaluation of general-purpose agents, scoring highly on real-world task completion and leading in cost efficiency. Its performance highlights how agent harness engineering—not just the underlying model—drives reliability, affordability, and safety in complex workflows.
- IDC evaluated nearly 100 private office tasks, including cleaning 3,755 rows of data and turning roughly 30,000 Chinese characters into an 11-slide presentation; TeleAgent scored 3.49 on routine tasks and 3.36 on complex tasks.
- Its agent harness uses tiered context pruning and compression, a context window exceeding 400K, and autoDream long-term memory to support lengthy and cross-session work.
- A ModelRouter assigns requests to lightweight, balanced, or flagship models, reportedly reducing inference costs by about 40%; the system also uses prefill/decode separation, KV caching, and load-aware scheduling.
- Enterprise safeguards include skill scanning, separated short- and long-term memory, restricted file access, high-risk command monitoring, and isolated execution; TeleAgent reached nearly 1.2 million users shortly after its July 2026 public launch.
Sources (1)
央企做了个通用Agent，直接杀进IDC实测前三！
TL;DR - China Telecom’s enterprise-focused TeleAgent ranked third in IDC China’s evaluation of general-purpose agents, scoring highly on real-world task completion and leading in cost efficiency. Its performance highlights how agent harness engineering—not just the underlying model—drives reliability, affordability, and safety in complex workflows.
- IDC evaluated nearly 100 private office tasks, including cleaning 3,755 rows of data and turning roughly 30,000 Chinese characters into an 11-slide presentation; TeleAgent scored 3.49 on routine tasks and 3.36 on complex tasks.
- Its agent harness uses tiered context pruning and compression, a context window exceeding 400K, and autoDream long-term memory to support lengthy and cross-session work.
- A ModelRouter assigns requests to lightweight, balanced, or flagship models, reportedly reducing inference costs by about 40%; the system also uses prefill/decode separation, KV caching, and load-aware scheduling.
- Enterprise safeguards include skill scanning, separated short- and long-term memory, restricted file access, high-risk command monitoring, and isolated execution; TeleAgent reached nearly 1.2 million users shortly after its July 2026 public launch.