Jensen Huang: Even AGI Won't Auto-Boost Company Productivity — You Need an Agent Harness
大模型之路 · wechat · 2026-09-12
At a G20 innovation ministers meeting, Jensen Huang pushed back on the idea that plugging in AI automatically raises productivity: "That's just not going to happen at all." Strong models still need a full engineering environment around them, which he calls the Agent Harness — an exoskeleton for the model's brain.
Key points
- He uses a new-hire analogy: even a MIT/Stanford PhD doesn't know where the code lives on day one. Models need context, purpose, relevance, and access to do real work.
- The Harness boils down to three blocks: feed the model business context (docs, tickets), define goals and boundaries via prompts/config, and grant whitelisted tool/data access — plus audit logging on top.
- His implicit thesis: models are commodities that get cheaper and stronger every six months; the deterministic, accumulable Harness is the actual moat.
Author's take
Don't treat the Harness as heavy bespoke engineering: provision context/purpose/access, add human confirmation for critical actions, and log every step for replay. Keep iterating it as business changes — good guardrails draw boundaries rather than lock agents down. This deployment paradigm complements Sutton's "era of experience" training paradigm.
More from coding & agent
- Devs split on GPT-6 Astra: fast and token-cheap but code feels 'alien'; costly stack proposed — beffjezos · 2026-09-12
- Pwnisher winner tests full GPT Astra x Blender x Dreamina 3D animation pipeline — aziz4ai · 2026-09-12
- 1,740 runs show token compression tools barely cut the final bill — Bartaseth · 2026-09-12
- Running 100-200 agents daily: disk space is now the bottleneck, not compute — vincent_koc · 2026-09-12
- Matrix × Mario vibe-coded with a 35B model on 6GB VRAM, full prompt and setup included — BrianScottGregory · 2026-09-12
- Codex team resets everyone's limits over a bug affecting just 4-5K of 20M+ users — i_dg23 · 2026-09-12