Event Overview
At the CCF YOCSEF Hangzhou Tech Forum on June 7, organized by the China Computer Federation with backing from Alibaba ATH-Qwen, Zhu Da—head of Qwen's C-end MOS (Model-Service) Lab—delivered a talk titled "Qwen C-End Agent Harness Thinking and Practice."
Zhu Da positioned his team as 'pragmatists': rather than chasing the theoretical upper limit of model capability, they focus on concrete problems like "how to serve over 300 million monthly active users with Qwen models." The centerpiece of the talk was the General Complex-Task Agent, launched in January 2026 as the 'capsule' entry above the Qwen App chat box. He condensed its engineering methodology into four characters: 多快好省 (Multi-Fast-Good-Cheap).
The hard numbers reported: execution time reduced to 1/3 of initial baseline, Token consumption at 1/10 or less of overseas peers. On a super-app with over 100 million MAU, this is an unambiguous engineering scorecard.
In-Depth Analysis
1. 'Multi-Fast-Good-Cheap' is an e-commerce-style framework for C-end Agent engineering.
Zhu Da's breakdown is remarkably e-commerce-like:
- Multi: Support diverse task types—information gathering, research analysis, life services, office collaboration, code development. A single general Agent covers all types, rather than building N vertical workflows.
- Fast: Execution time must match delivery quality—a 5-minute / 10-minute / half-hour task should produce stepwise-improved outputs. Key techniques: task parallelization + 'path solidification' (first encounters with a task need heavy reflection; accumulated experience then freezes execution paths into code).
- Good: The hardest industry problem is not 'how to do well' but 'how to define good.' Collecting all coin varieties across Chinese dynasties from Qin to Qing has no clear 'completion' criterion—ordinary Agents stop after finding a dozen. Optimization requires search-paradigm tuning + context management to 'force' the model to continue.
- Cheap: The product is free, so cost control is a lifeline. Reduce model calls, let the model write code instead of repeating inference, solidify paths, optimize caching.
2. 'Proactive Service' is the most noteworthy paradigm upgrade for C-end Agents in H2 2026.
Zhu Da sketched a real personal-assistant persona: not 'I constantly ask my assistant,' but an assistant that knows everything about you and proactively serves and pushes.
The architecture is four components: User Memory (user profile) + Environment (context awareness) + Task System (task management) + Assistant (unified touchpoint).
The biggest challenge is 'EQ'—Zhu Da's example was very specific: a user once mentioned a certain medical condition, and when recommending a restaurant the AI noted 'this one is too spicy, you mentioned before…' Sounds reasonable, but the user perceives it as the AI over-exploiting privacy. Say too much and it's intrusive; say too little and it's invisible. Proactive service's demand on model intelligence, in the dimension of 'sense of proportion,' currently exceeds what base models can solve. Zhu Da's judgment: this may need to be built into the base model itself, not patched via post-training or PE.
3. The evolution 'Harness Engineering → AIWare Engineering,' justified through computer-history analogy.
Zhu Da used a 'computer-development history' metaphor to divide Agent engineering into four stages:
| Year | Computing Milestone | Agent Engineering Stage | |---|---|---| | 1946 | ENIAC birth | Large model (raw compute) | | 1949 | Assembly language | Prompt Engineering | | 1968 | NATO Software Crisis Conference ('software engineering' coined) | AIWare Engineering | | 1969 | Unix OS | Agent framework / Harness | | 1972 | C language | Context Engineering |
Key insight: 1968 was the true turning point—back then CPU, OS, and C language had all appeared; everyone thought software problems were solved, only to discover the real difficulty was 'software is not a code problem but a human problem—expectations, project management, team organization.' Translated to the AI era: the next step after Harness Engineering is AIWare Engineering—how to truly integrate people and AI.
4. 'Low-power, good-enough' is the most counter-consensus judgment in this talk.
Zhu Da cited human evolution: humans evolved language and increased brain capacity 70,000 years ago; since then—writing, paper, printing, radio, the internet—human brain capacity has not increased again. Subsequent progress was essentially optimizing 'context engineering'—information storage, transmission, faster exchange, more efficient collaboration.
Applied to AI, his core view: the model can be more accurate, but is the ROI worth it? Is the energy demand reasonable? Perhaps using lower power + better Harness methods can roughly approach higher-model effects—and for societal development overall, this is more appropriate. Base model + Harness, ultimately forming a balanced state—not endlessly pursuing single-dimension extremes.
Why It Matters
This is a rare C-end Agent retrospective with both 'macro vision' and 'hard engineering numbers.' Three signals stand out:
Signal 1: 1/3 time + 1/10 Tokens—C-end Agent cost curves have been pushed to the limit by Chinese tech giants. This is a critical benchmark for global C-end Agent commercialization over the next 12–18 months.
Signal 2: 'Proactive service' will become the main battleground for C-end Agents in H2 2026. By pushing this line into product form on an app with 300M+ MAU, Qwen forces every competitor to directly answer 'how will you proactively push.'
Signal 3: 'AIWare Engineering' is becoming a new buzzword. From Prompt → Context → Harness → AIWare, this engineering-evolution path will be cited repeatedly in upcoming YC talks, podcasts, and founder camps over the next 6 months—it is a concept mine.
Risks and Open Questions
First, Zhu Da's observation that 'many prior scaffolds become unnecessary as base models improve' carries cautionary value—many current Harness Engineering patches will be eaten by the next generation of base models. This means Harness Engineering's moat may be shorter than imagined.
Second, the hardest problem—'EQ'—has not seen a systematic solution from any model or team yet. Qwen's current 'proactive service' is likely still in a 'fine-grained probing' phase, not a complete capability breakthrough. Watching Qwen App's actual behavior over the next 3–6 months will be more informative than re-reading this talk.
Third, 'low-power good-enough' is an engineering philosophy, not a business strategy. If other majors (especially the OpenAI/Anthropic camp) keep 'higher intelligence, stronger base model' as their competitive main line, whether Qwen's 'low-power philosophy' can withstand the China-US base-model generation gap needs time to validate.
---
Source: Alibaba ATH-Qwen MOS Lab head Zhu Da's keynote at the CCF YOCSEF Hangzhou Tech Forum (June 7, 2026).