Zhihu Frontier Weekly｜Harness Engineering Emerges, Sora Shuts Down, and RISC-V Heats Up
From prompt-level tricks to system-level control, AI is entering the era of engineered agents
Welcome to Zhihu Frontier, your window into the hottest AI convos from China’s knowledge platform.
This week, the conversation shifts from models to systems.
Harness Engineering is emerging as a new paradigm for building reliable AI agents. Meanwhile, RISC-V is gaining real traction in the AI hardware stack, and decisions from OpenAI and Kimi reveal deeper trade-offs between capability, cost, and usability.
🧩 Harness Engineering Emerges as the Missing Layer for Reliable AI Agents
Read the full post: View original
Zhihu contributor @李进锋:
If you’ve used tools like Claude Code, Cursor, or similar AI coding tools, you’ve likely seen this: writing scripts or fixing bugs works fine—but asking for a full web app often results in messy, unmaintainable output. Context runs out, or the result becomes impossible to reproduce or audit.
This isn’t a user issue—it’s an industry-wide limitation.
Anthropic and OpenAI are both exploring Harness Engineering: a framework of structured constraints and feedback loops that make AI outputs controllable, auditable, and iterative.
As OpenAI puts it:
“What Codex can’t see doesn’t exist.”
If rules and context live only in human heads or scattered tools, they are invisible to agents. Harness Engineering encodes them into machine-readable structure—turning implicit knowledge into enforceable systems.
In essence, it acts like a “harness” for AI—guiding behavior instead of letting it run freely.
🔁 From Prompt to Harness: AI Engineering Shifts Toward System-Level Control
Explore the original discussion: Zhihu column
Zhihu contributor @Weyne Chen:
AI engineering has evolved through three stages:
Prompt Engineering → Context Engineering → Harness EngineeringPrompts help with single-turn responses. But for long-running tasks, they fall short.
Failures often aren’t about intelligence—they’re about missing structure: what to look at, what tools to use, how to recover from errors.
These three layers form one system:
Prompt → how tasks are defined
Context → what the model can access
Harness → how execution is structured
As goals shift from answering questions to completing workflows, engineering focus moves outward—toward environment design, feedback loops, and system constraints.
⚙️ Alibaba’s Xuantie C950 Pushes RISC-V to Parity with x86 and ARM
See the Zhihu thread: Discussion link
Zhihu contributor @包云岗:
Alibaba’s Xuantie C950 reaches SPEC2006INT 22/GHz, surpassing Tenstorrent’s Ascalon and becoming the strongest publicly known RISC-V core.
Its performance is now comparable to top x86 and ARM CPUs. Chips based on it are expected in 1–2 years.
Major companies like Google, Meta, Microsoft, and Tesla are already adopting RISC-V for AI chips. Meanwhile, global standards are forming: vector extensions are unified, and matrix/tensor standards are underway.
This could enable a full “RISC-V + AI” ecosystem at both hardware and software levels.
🤝 Cursor Responds to “Kimi Wrapper” Controversy, Confirms Authorized Integration
Join the discussion here: Zhihu thread
Zhihu contributor @Kimi:
Thanks for the attention.
Congrats to Cursor on launching Composer 2.
We’re glad to see Kimi K2.5 serving as part of its foundation. Its integration via continued pretraining and RL optimization aligns with the open ecosystem we aim to support.
(Cursor accesses Kimi K2.5 through Fireworks AI as part of an authorized partnership.)
🎬 OpenAI Shuts Down Sora, Signaling Product Consolidation Under Compute Pressure
More context: Full answer
Zhihu contributor @赵泠:
Sora’s shutdown is expected. It consumed massive compute resources while failing to retain users.
This isn’t about geopolitics—it’s about product viability.
While Sora 2 improves realism and control, these aren’t standout breakthroughs.
The real issues were product-level: weak user guidance, lack of community features, and poor engagement design.
📚 Recommended Reading
Token translation discussion: View thread
Deep dive on Superpowers: Read more
Full Superpowers guide: Full article
NeurIPS controversy update: Details
WeChat ClawBot plugin launch: Discussion
📬 That’s all for this week’s AI round-up from Zhihu Frontier.
👉 Subscribe to never miss an update: zhihufrontier.substack.com