Overview
On June 7, 2026, at a CCF YOCSEF Hangzhou technical forum supported by Alibaba's ATH-Qwen business group, Zhu Da, head of Qwen's consumer-side (C-end) MOS Lab, gave a talk titled *Reflections and Practices on Qwen's C-end Agent Harness*.
Zhu described his team as "pragmatists" — rather than chasing theoretical limits of model capability, they focus on concrete problems like "how to serve 300M+ monthly active users with Qwen models." The centerpiece is the General Complex Task Agent (the "capsule" entry above the Qwen App chat box), officially launched in January 2026. His engineering methodology condenses into four characters: "duo, kuai, hao, sheng" — more, faster, better, cheaper.
The hard numbers: execution time reduced to 1/3 of the initial version, token consumption only 1/10 or less of comparable overseas products.
The "More, Faster, Better, Cheaper" Framework
- More: Support information gathering, research analysis, lifestyle services, office collaboration, and coding — one general agent covering all task types rather than N vertical workflows.
- Faster: Execution time should map to delivery quality — a 5-minute, 10-minute, or half-hour run should yield stepwise improvements. Key methods: task parallelization plus "path consolidation" (first-time tasks require heavy reflection; accumulated experience lets execution paths be hardened into code).
- Better: The hardest part isn't doing well, it's *defining* well. For open-ended tasks like collecting all coin types from China's Qin to Qing dynasties, ordinary agents stop after a dozen; search-paradigm optimization and context management must "push" the model to continue.
- Cheaper: The product is free, so cost control is the lifeline — fewer model calls, letting the model write code to replace repetitive reasoning, path consolidation, cache optimization.
- Many current harness scaffolding patches will be absorbed by next-generation base models — harness moats may be shorter than expected.
- "EQ" has no systematic solution from any team yet; Qwen's proactive service is likely still in fine-grained probing rather than a full capability breakthrough. Watching the Qwen App's actual behavior over 3–6 months will be more informative than the talk itself.
- "Low power, good enough" is an engineering philosophy, not a business strategy. If OpenAI/Anthropic camps keep pursuing higher intelligence as the competitive mainline, whether Qwen's philosophy withstands the base-model gap remains to be seen.
The distinctive point: this methodology is designed entirely around the constraint of "serving 300M+ MAU," not derived backward from model capability. It's product-driven engineering — the reverse of most LLM companies' path.
"Proactive Service": The Next Paradigm
Zhu outlined a true personal-assistant vision: not "I keep asking my assistant," but an assistant that knows everything about you and serves/pushes proactively.
The architecture has four components: User Memory + Environment + Task System + Assistant.
The hardest problem is "EQ." Example: if a user previously mentioned a disease, an AI recommending restaurants that says "this one is too spicy, given your condition..." sounds reasonable but feels like privacy overreach. Too much feels intrusive; too little feels absent. Zhu's judgment: this may need to exist at the base-model level, not something post-training or prompt engineering can patch.
From Harness Engineering to AIWare Engineering
Zhu mapped agent engineering onto computing history:
| Year | Computing history | Agent engineering stage | |---|---|---| | 1946 | ENIAC | Large models (raw compute) | | 1949 | Assembly language | Prompt Engineering | | 1968 | NATO software crisis conference ("software engineering" coined) | AIWare Engineering | | 1969 | Unix OS | Agent frameworks / Harness | | 1972 | C language | Context Engineering |
The key insight: 1968 was the real turning point — CPU, OS, and C existed, and people thought software was solved, only to find the real difficulty was that software solves human problems — expectations, project management, team organization, not just code. Transposed to AI: the next step after Harness Engineering is AIWare Engineering — truly integrating humans and AI.
The Contrarian Call: Low Power, "Good Enough"
Zhu invoked human evolution: language and brain growth ~70,000 years ago, then writing, paper, printing, radio, internet — with no further brain-capacity increase. Later progress was essentially optimizing "context engineering": storing, transmitting, and exchanging information more efficiently.
Applied to AI: models can be more precise, but is the ROI worth it? Are the energy and cost demands reasonable? Lower power with better harness methods may roughly match a bigger model — a more suitable outcome for society overall. Base model + harness should reach a reasonable equilibrium, not endless pursuit of single-dimension extremes.
Why It Matters
1. 1/3 time + 1/10 token: consumer-agent cost curves have been pushed to the limit by Chinese majors — a key reference for global consumer-agent commercialization over the next 12–18 months. 2. Proactive service will be the main battlefield of H2 2026; with 300M+ MAU, Qwen forces competitors to answer "how will you push proactively?" 3. "AIWare Engineering" is poised to become a hot term — the Prompt → Context → Harness → AIWare evolution path is a conceptual mine for the next six months.