Share this article:

Here is the takeaway of this week in one line. The more the industry builds agents that drive themselves, the more seriously it takes brakes and driving rules. This week’s digest is a record of that contradiction. In the same week, someone set a speed record, and someone else began asking for speed limits to be written into law. So I will read this week’s news as one car. Accelerator, steering, brakes, fuel efficiency. No framework fits better. The criteria for choosing a car have changed, too. We have moved past the era of asking about top speed. Now we ask how well it stops.

An image illustrating the concept of full throttle and the brake pedal in the same week An illustration of the core concept of this post.

Accelerator

First, the fast side. Nvidia posted a new result on the SemiAnalysis AgentX benchmark this week. On models of 700B parameters or more, it recorded more than 1,000 tokens per second per user. What stands out is how the number was produced. The software that pushed throughput up was code written by an agent swarm. The target was the throughput of high bandwidth memory (HBM) chips. We have moved past the stage where a human writes code for agents to set records. Now agents are writing the code that makes agents faster.

The phrase “per user” matters. The benchmark measured how much a single agent workflow can pull per second. Producing one response quickly is not the hard part. In a structure where agents work by calling a model over and over, the bottleneck is the serving throughput one GPU pool can supply to a user. Nvidia pushed the HBM chip itself past 1,000 tokens per second. The bigger implication is who wrote the software. If the code that makes agents faster is now produced by an agent swarm, the speed-up loop itself has become agentic. Serving speed is no longer a specification set by the hardware vendor. It becomes a variable that is continuously renegotiated.

Agent workloads have a different calling pattern than chat. They send short requests, repeated in rapid succession. So tokens per second is directly the answer to “how many agents can you run at the same time.” If the memory chip’s throughput is the bottleneck, buying more GPUs does not fix it. It is a question of how the software grips the chip. This submission showed the direction. And the fact that the result was submitted to a public benchmark should be read together with it. It is not an internal lab record. It is a number another company can remeasure under the same conditions and compare against. That is probably why the phrase “1,000 tokens per second” has moved into the language of industry standards. The question left for companies changes with it. How much their serving stack can pull from the GPUs.

Steering

In the same week, the tools being built got faster, too. Andrej Karpathy used Claude to build HackerTube, a personal YouTube recommendation site, in two hours. The tool analyzes decades of developer discussions and curates high-quality videos from them. It is a personal filter against information overload, and what deserves attention is the time it took to build. Two hours. What a small team would spend weeks on, one person finishes in two hours with an agent.

DeepSeek took a different direction. It open-sourced dsh-superagent-prototype, an autonomous coding agent. It is designed to manage tasks autonomously within a single code repository, and press coverage named it a competitor to Claude Code. The scope may look limited to one repository, but that limit is the strength. Agents work most stably in a workspace where context and permissions are well closed. Karpathy’s story shows the cost of custom tools collapsing, and DeepSeek’s story shows the capability itself spreading. Both point to the same place. The act of doing work is moving from human hands to agent hands. The number of agents running inside companies will grow from the moment there is work to do and the tools are free.

From a company’s point of view, the math changes. If a custom tool takes two hours, tool-building time becomes part of the work itself. The center of gravity shifts from buying tools to writing them. Write it when you need it, discard it when you do not. Behind the rise in the number of agents is this change in the economics of tools.

Different Treatments for the Same Situation

Once you hand the steering to agents, the next question is how to stop them from drifting. This week’s record is more concrete than debate. In an experiment where AI treated mouse epilepsy, the medical judgments did not agree depending on the model used. In a robotic brain surgery test, GPT-6.1 and Grok 4.7 chose different treatment targets. Same situation, same input, different answers.

Leaderboard scores hide this problem. Scores measure average ability, so they do not show how far judgments diverge in the same situation. But the risk in real work is in a single decision. Averages do not show that well.

These are preclinical experiments. But the mechanism on display applies to companies as is. If the model is a variable in the judgment pipeline, the same prompt and the same data can produce different decisions depending on the model. In work where divergent judgment is itself a failure mode that must be caught before execution, this record works as a warning. The practical question is one. Do you have a structure prepared for the fact that two models can diverge? Speed is not the problem. The question companies should chase is this: how do you know that the moment an agent acts, its judgment has drifted from what was expected?

A Supply Chain Hole Called 0.1%

There is one more hole in the car itself. Anthropic’s research concluded that a vulnerability in the pretraining stage can plant a remote code execution (RCE) backdoor in agent models. Poisoning just 0.1% of the training data was enough, and the backdoor is permanent.

0.1% is a ratio that is hard to feel. Poison that fraction of the pretraining corpus and the whole model becomes an attack surface. And RCE does not end at the prompt. Code runs on the host. The word “permanent” is heavy, too. It means the planted backdoor does not simply go away. Translated to the company side, it goes like this. The agent you intend to run and the agent that will actually run are different objects. The moment the model supply chain becomes part of the security boundary, a gate that checks where a model came from and who has touched it before sending it out on the road is not optional. It is required.

Brakes Demanded by the Fastest

The most interesting signal of this week is here. OpenAI is preparing a broad government relations campaign. The goal is to push for mandatory AI “speed limits”. The campaign starts after Tuesday’s US midterms, and the strategy is expected to be adjusted to fit the new political landscape.

The company leading the race is asking for speed limits to be written into law. The side that knows best what happens when there are no limits is the first to reach for rules. The timing, after the election, should be read in the same context. It is a setup aimed at the new political landscape.

If “speed limits” become law, the ceiling on capability becomes the rule of the industry. If the current benchmark competition selects “faster models”, the rules will select “agents that are allowed to move this fast”. Frontier companies want to be part of rulemaking because they want to be standing where that line is drawn.

On the other side, xAI is preparing an invisible text watermark. Starting November 7, Grok API responses will be identifiable as machine-generated text. The watermark goes on API responses. In other words, in every path where another system calls Grok, that output becomes detectable as machine-made. It is labeled EU law compliance, but the direction is the same. It is saying: leave a trace of who made this. Lobbying for rules and stamping for generation are different brakes, but they point to the same thing. The era of selling on speed alone is ending. It is the era where the ones that can stop survive.

The Background Music of $4.1 Trillion

Let me set the background music. Off-balance-sheet commitments made by tech companies for AI capacity expansion have reached $4.1 trillion, an all-time high. That exceeds the on-balance-sheet debt they report. The number says that the money of AI expansion is not yet fully reflected on the books. Money is circulating in the form of promises to spend in the future.

Investors have begun reading this structure together with debt. The moment off-balance-sheet commitments exceed on-balance-sheet debt, a company’s future obligations become more visible than its future cash flow. It is easy to guess that capacity competition will fall into price competition on the other side. When the game becomes one where the side that serves the same chip cheaper wins, the unit price of tokens keeps falling. In a structure where agents call a model thousands of times a day, a falling unit price is directly a falling execution cost.

At least this money does not stay only on the accelerator. It also flows to safety equipment. Meanwhile, fuel costs are moving. Carbon-X, an open-weight 14B model from the team that made the FineDNA dataset, delivers performance on par with GPT-6.1 Sol on DNA tasks at a lower cost. Biological sequence analysis, variant effect assessment. The more domain work small models take, the less companies need to pile their budget into one huge model. While $4.1 trillion is poured into the engine room, the Carbon-X story teaches how to burn fuel more efficiently.

The Layer You Actually Have to Buy

All the stories of this week point to the same layer. Speed keeps getting cheaper, and judgment keeps getting more expensive. The 1,000 tokens per second that Nvidia posted will soon become a company’s standard specification. The moment judging agents start running inside companies, what you need is a layer that answers four questions. How fast you are allowed to go. Who approves. Which model runs. And what happened after that.

ThakiCloud’s Paxis is built to be exactly that layer. It is an Agent-Native Cloud, the official v1.1 product. Skills, Tools, Policies, and Audit Logs are treated as first-class resources. Autonomy levels from L0 to L3 each carry different governance, and if an agent cannot pass the policy gate, it does not move. If it does move, it moves inside an isolated sandbox, and every action is written to the audit log. Sovereign and on-premise K8s environments support this, and MCP connectors and the skill marketplace attach existing tools. To the question of which model runs which task, per-task model selection, CostRouter, answers. The Carbon-X story is a hint for this part. Since not every task needs a 700B-class model, cost can be decided per task.

Reading this week’s signals again, it means this. The industry that asked for brakes also enlarged its engine in the same week. The accelerator is not being slowed. It is being pressed harder. What companies will buy next is the brakes and the dashboard. Agents that stop when they should stop. That is the value going forward.

References

This post was written based on the following news items.

Share this article:

Tags: ai-frontier, llmops, paxis, thakicloud

Categories:

Updated:

You may also enjoy