The AI slowdown nobody can agree on

Some frontier lab leaders want room to slow the frontier, but the governments racing to lead it are not keen on touching the brakes. Meanwhile, two live deployments show where agents are helping and where extra agents made the work worse.

The AI safety argument has moved from whether frontier systems could become dangerous to who would accept slowing them down. Lab leaders have floated pacing plans, independent evaluations and tighter controls on the most capable models. The political response is running into a harder problem because neither Washington nor Beijing wants to give the other side a cleaner run at the frontier.

The US has not formally rejected one shared plan. The resistance is broader. The Trump administration has played down AI risk, Congress is unlikely to build a regime quickly, and coordination between leading labs could hit antitrust questions. China has pushed back on the slowdown framing and on measures that look like technological containment, including tighter chip restrictions.

That leaves the people asking for brakes trying to design them while the race continues. Anthropic’s proposals include independent evaluators, stronger model security and government power to block a deployment judged catastrophically dangerous. Each step still depends on political coordination that does not exist yet.

Customer support agents are easiest to judge when the job is narrow, repetitive and easy to hand back to a person. TSA’s new AI agent, Ace, has now reached that kind of scale. Salesforce says it handles about 100,000 routine traveller conversations each month and resolves 96% without human escalation.

Ace answers questions such as how to pack liquids or pass through a checkpoint with a medical device. It runs on Salesforce Government Cloud and uses TSA guidance to return an answer with the right context. More complex requests still move to staff through connected contact-centre and regional systems.

The 96% rate covers routine enquiries, not every TSA conversation, and it measures resolution without escalation rather than accuracy or satisfaction. Salesforce has not supplied an independent audit or detailed measurement window.

Founders context: A narrow support queue gives you a cleaner first agent project than an open-ended assistant. Track escalation rate, wrong answers and repeat contacts separately before claiming success.

More agents can create more handovers, and Polylane found that the handovers were eating the evidence. Its autofix system once used a triage agent, a coordinator, as many as 15 investigation agents and a coding agent. A confirmed issue could pass through 18 agents before the system opened a pull request.

Each specialist saw a summary rather than the full trail. Investigators repeated work, the coding agent received a plan without the original evidence, and individual stages could pass their tests while the full run failed between them. Polylane replaced the chain with one agent that handles triage, investigation, root-cause analysis and the pull request in a single trace.

The company reports sharp early changes after the 3 September switch:

Those numbers are self-reported and cover one company’s autofix system. They do not prove that every multi-agent design is wrong.

Founders context: Before adding another agent, inspect what gets lost at each handover. One longer trace can beat a neat org chart of specialists.

🍎 Apple begins its Siri AI rollout across its 2027 software releases, pushing its delayed assistant upgrade into customers’ hands.

📦 Box connects enterprise content to ChatGPT so staff can work with authorised files without moving copies into a separate store.

🧠 Google and HHMI open a male fruit-fly brain map for researchers and developers to explore with AI tools.

🛡️ MIT’s HardFlow method aims to keep generated outputs inside strict constraints for safety-critical systems.

💬 404 Media reports human review inside ChatGPT, raising questions about which staff can see sensitive conversations and under what controls.