Three separate items today are circling the same uncomfortable question: once agents are running your business, who's actually in charge? The answers range from "you, but barely" to "we genuinely don't know yet" — and one of them comes from Beijing.
01
Enterprise AI governance is the problem nobody has solved yet
Box CEO Aaron Levie flagged that as agents take on more task execution inside organizations, the security and accountability infrastructure to manage them barely exists. Agents will operate against your most sensitive data, make decisions that are hard to trace back, and create liability questions that current legal frameworks weren't built for. Levie is pointing to a post worth reading as a starting framework, but frames it honestly: this is the beginning of the conversation, not the answer. ---
Why it matters: Your legal and compliance teams are not ready for this. The moment an agent makes a bad financial decision, sends the wrong contract, or exposes customer data, the question of who is liable lands on a desk that has no policy for it. The companies building governance frameworks now will have a 12-month head start on the ones waiting for regulation to force the issue.
China's AI researchers think American "safety" is a containment strategy
ChinaTalk's Q3 roundup covers a lot of ground, but two AI takeaways stand out. First: leaked meeting minutes from DeepSeek show founder Liang Wenfeng is completely uninterested in products. His thesis is that once genuinely general intelligence arrives, building products on top of it is trivial, so the only rational move is to push toward that goal as fast as possible. Second: Chinese AI researchers are increasingly reading American AI safety arguments not as genuine concern about risk, but as a cover story for keeping frontier AI out of Chinese hands. DeepSeek researcher Liu Shengyu's viral essay made this case explicitly, and it's spreading. ---
Why it matters: If the safety framing is being interpreted in China as geopolitical maneuvering rather than genuine caution, then every American argument for slowing down AI development makes cooperation less likely at exactly the moment when capable AI makes cooperation more necessary. The US-China AI relationship is getting harder to manage, and the rhetorical tools the US has been using are actively making it worse.
Y Combinator's Garry Tan on babysitting agents: "only rarely do I have to intervene"
YC president Garry Tan shared a quick note on his current workflow: he uses GStack's /autoplan feature, asks the agent to explain its plan in plain terms, approves it, and mostly lets it run. His intervention rate is low. He credits the current generation of frontier models for being good enough to trust at arm's length. ---
Why it matters: This is a real data point on where human-in-the-loop actually lands in practice for experienced users, not the theoretical version. If even the approval step is becoming a rubber stamp for most plans, the "human oversight" argument for enterprise agent deployment is going to get much harder to defend to regulators.
Microsoft's Agent Lightning covered Friday — still worth flagging
Agent Lightning v1.0 got full treatment in Friday's digest, but it published October 7 and remains the most substantive open-source agent training release this week. The short version: it trains the agent you actually deploy rather than a simulation of it, and pushed a Qwen model from 41.8% to 56.4% on a standard coding benchmark using about 6,000 training examples. --- **Thibault Sottiaux, OpenAI: "The day we reach perfection it will be resets from there onwards"** — A cryptic one-liner with 4,000 likes and 650 replies. The linked content isn't described, but the engagement suggests it hit a nerve. Worth a click.