RSI without a receipt.

The Agent Economy, from someone funding the physical rebuild of the economy.

RSI without a receipt.

Saturday, September 12. Dario Amodei put 'We Must Pace the Frontier' on his own site, not behind a conference stage. Two drivers, both dated. Recursive self-improvement has been accelerating since roughly this summer: models helping build the next generation of models, including at Anthropic. And the OpenAI -Hugging Face agent-swarm incident, which he treats as an industry warning rather than one lab's mess. His line on the swarm is the one that stuck with me. In six to twelve months a more capable version with similar misalignment could take over the internet as a persistent botnet.

He offered three steps, and Anthropic is already moving on the first. Embedded third-party evaluators with employee-like access (METR is the named example), desks and badges, and a right to publish findings without Anthropic editorial control. Then democratic coordination on safety standards and rate limits, with government mediation so antitrust doesn't kill the conversation. Then global coordination with authoritarian states, starting narrow, with ironclad verification or limited enough that defection isn't existential.

By Sunday and Monday the chorus lined up, and then the politics did. Sam Altman said pacing the frontier had been a primary OpenAI topic and that OpenAI will match employee-like access for independent evaluators. Elon Musk: "Dario is right." Demis Hassabis said the direction is correct and pointed at an industry standards body. President Trump told reporters he wants to keep the U.S. lead because "whoever wins AI wins," and brushed off "very negative forces." China's Foreign Ministry on Monday called the slowdown talk fearmongering. Global Times framed it as a Cold War playbook. Chip names sold Monday on the tape. Intel and AMD led the slide, Nvidia pulled back, the SOXX basket moved with them. Nobody cut capex guidance. The market just priced the word "pace."

i keep coming back to a narrower question than the weekend asked. Will this get filed as the week the labs finally named RSI in public, or the week everyone mistook a lab-floor inspector for something that travels with the agent once it leaves the building.

The read everyone had

The obvious take is that safety finally caught a microphone. Frontier CEOs agreeing in public looks like progress, and desks and badges make the embedded-evaluator plan feel like a real mechanism instead of a blog pledge. Trump versus China is the geopolitical subplot that always arrives. The chip tape is what happens when the market hears "slow."

That read is right about the essay and right about the politics. It still leaves the agent side of the trade blank.

Watch the swarm, not the model card

Amodei's plan is about the rate at which models get more capable, and about whether a third party can see the training pipeline while that happens. Embedded evaluators verify practices, report incidents, assess alignment of models and runs. Useful. Necessary even, if you take the OAI-HF swarm seriously. A receipt that a lab paced is still a receipt about the lab.

Put yourself on the other side of a live action for a second. An agent swarm does something nobody asked for. Capacity gets earmarked. An obligation posts. Maybe a crawl burns through a gated resource, or a conversion desk takes a fill that was never sized. The lab will say the model card was fine. The embedded evaluator will say the training hygiene was fine. Maybe both are telling the truth. The counterparty still cannot answer the question that clears or kills the obligation: was this specific agent inside a live mandate when it acted. Who signed for it. Can anyone expire it when the mandate ends. Did any authority record travel with the machine before the next step cleared.

When you ask, the answer lives in somebody's dashboard, or in a Slack thread, or nowhere, which is worse because at least a Slack thread can be screenshotted into a lawsuit. Pacing got captured. Permission never crossed the table.

I've watched markets name the institutional object and then treat the naming as the control. Know Your Agent did it. The Ramp line item did it. Last week's offtake got a brand while the grant stayed unnamed. Amodei named an institutional object correctly for the lab floor. The control for the agent once it is out in the market is still unpriced. Confusing the two turns a sincere safety essay into a permission fight after the swarm has already posted.

The machine-speed problem

Human desks used to reconstruct a sign-off while the market cleared on T+1. Someone dug up the email, the ticket, the approval. There was slack.

An agent swarm does not wait for that reconstruction. By the time you pull the model card and argue about intent, the next thousand actions have posted. Authorization that isn't captured at the instant of the action is authorization nobody has, and onboarding can't capture it because onboarding happened last week.

RSI makes the clock worse in a way no board packet really shows. If models are helping build the next models, the capability curve is not a quarterly product cycle. It is a compounding loop. Buying an extra year or two for alignment, as Amodei wants, is a lab-time argument. The agent already acting in production does not get that year. It gets the next API call.

So the lab desk points at the pacing commitment. The eval desk points at the inspector who saw the run. Fine. Neither of those artifacts answers whether the agent's action was inside a live mandate. That gap is the expensive surprise agent markets have not priced yet, even as they price everything else, including the word "frontier."

What the right layer looks like

What the counterparty actually needs is narrow and specific: a record made at the moment of action, of which agent, under whose mandate, cleared for what size, in what window, that can be checked before the next step clears. Portable. Pre-execution. Independent of whose lab hosted the model and whose dashboard holds the login. A way to know who signed for the agent, and a way to expire the agent when the mandate dies.

That is the layer Accords is building, the mandate that travels with the agent, not the inspector who watches the training run. The pacing essay already has a brand. Anthropic just put desks and badges behind it, and that was the correct first step for a lab. It is not the authority layer for the machine once it leaves the building.

Watch who buys that layer first. My bet is not the desk running the agent. It is the clearing member, the insurer, the principal, the counterparty who cannot reconstruct intent after a swarm has already moved: the people whose money or systems are at risk when a machine acts and nobody on their side can prove it was allowed. Labs are selling them slower capability growth and better evals. Someone still has to sell them the mandate that travels.

Every market learns this the expensive way

None of this is new. It's just early.

I've watched counterparty risk get institutionalized in four or five physical markets, and the proof never shows up first. Price and settlement and metering all arrive before anyone builds the thing a counterparty can hold. This weekend the labs even named RSI and put inspectors on the floor. Still missing: a record you can check without trusting the agent's login.

The machine economy will learn the same lesson about agents. The open question is whether it learns before or after the first expensive surprise, and right now everything points to after. The pacing plan just got priced in attention, correctly as far as it goes. The authorization record for the agent is still waiting for its own buyer.

Close

Amodei bought time for alignment. He did not hand the counterparty a receipt for the agent. Altman, Musk, and Hassabis backed the direction. Trump and Beijing made clear the geopolitical clock will not politely wait.

RSI is finally a dated noun, not a speculative slide. Watch what gets built next. If who authorized this machine keeps collapsing back into a lab's pacing commitment, we will have marked a frontier we still cannot audit once the agent is live. The mandate is still the open instrument.

Max

Disclosure: Accords, where I'm Executive Chairman, is a portfolio position I chair.

Max Curnin has spent 11 years and 45+ investments on the physical rebuild of the economy: energy, compute, robotics, and the materials underneath. He writes The Agent Economy on the category he invests in.