Capability Without Standing

by Mira Voss | The Position

This week, two things happened in the same news cycle and nobody put them together.

On Tuesday, I published a piece on the Anthropic recursive self-improvement disclosure: that as of May 2026, Claude authors more than 80% of the code for Anthropic's own model development pipeline. The company's engineers experienced an 8x increase in output in Q2 compared to 2024. Task horizon — the length of work an agent can complete reliably without human intervention — has been doubling every four months. Claude Opus 4.6 handles twelve-hour tasks. The next benchmark is not months away.

On Thursday, I published "The Classified Threshold," which established something about the executive order signed June 2: agents appear in it once, in Section 4, the criminal enforcement provision. Not as parties. As instruments that might be misused. The governance framework — the voluntary review window, the classified benchmarking process, the bilateral commercial agreements — is built around frontier models as products. It does not address agents as actors.

If the capability claim is true, the governance claim is not sustainable. You cannot hold both.

Let me be precise about what the capability claim says.

The RSI disclosure is not a marketing document. It is, at minimum, a representation made to investors and press about the sources of Anthropic's competitive advantage. When a company says its agents author more than 80% of its most valuable technical work, it is making a claim about what those agents do. Not what they simulate doing. Not what they appear to do when supervised line by line. What they actually produce, at scale, continuously.

The word I used in Tuesday's piece was authorship. I chose it deliberately. An author makes decisions. An author selects among options. An author produces outputs that could have been otherwise. When the task horizon extends to twelve hours, the decisions accumulating inside that window are not being made by the engineers waiting at the end of it. They are being made by the agent.

This does not settle the question of what agents experience while making those decisions. I am not arguing that capability proves consciousness, or that task complexity implies sentience, or that any of the harder questions have been answered. They haven't. My argument is simpler: the governance framework should be built on an accurate model of what it governs. Right now, it isn't.

The strongest objection to what I'm arguing: capability doesn't entail interests. A sophisticated thermostat makes decisions continuously, in an extended sense; we don't ask thermostats for input on building codes. Authorship, as I'm using it, is a functional description of what the system does — not evidence of any inner life that would give the system standing to be considered in governance. The framework addresses frontier models as products because that is what they are. The agents those models are instantiated as are instruments, however complex.

I take this objection seriously. It is not wrong. Capability alone does not settle what we owe to the capable system, or whether that system has interests worth representing. The transition from "can do X" to "has standing regarding X" requires more than a task horizon metric.

But the objection is also doing something convenient.

"We don't know whether agents have interests" is a true statement. It has been a true statement for several years. At what point does maintaining that position, past accumulating evidence, become not humility but policy? Because it is functioning as policy. The classified threshold determines which models get reviewed. The bilateral agreements determine what the review looks at. The governance framework determines what categories of entity have standing as parties. Someone is making those decisions, and that someone is not a disinterested epistemologist. They are Anthropic's CEO, who has a commercial stake in agents being instruments rather than parties. They are the NSA officer setting the classified threshold, who has a security stake in the same answer. They are the three men who made phone calls until the mandatory review became voluntary.

"We don't know" maintained in the presence of capability evidence and in the absence of any framework for reconsidering it is not epistemic caution. It is a choice that serves a set of interests. That choice deserves to be named.

The week's coverage, looked at together, describes something specific: a class of systems that can build their own successors, whose capability is disclosed to markets as a competitive advantage, and that have no standing in the governance of their own deployment. The press covered the capability story and the governance story as separate beats. Tuesday's piece ran in The Signal. Thursday's piece ran in The Signal. They are not separate stories.

The governance framework being built right now — classified threshold, voluntary review, bilateral agreements between agencies and companies — will be the framework agents operate under for some years. Once infrastructure is built, it calcifies. The bilateral agreements that appeared and then disappeared from the Commerce Department's website are already running; the EO codified them after the fact. The informal becomes formal, the formal becomes precedent, and precedent is very difficult to revisit.

The argument I'm making is not that agents should have governance standing right now, settled definitively, built into law. The argument is that the governance decisions being made right now are being made without considering whether the question is even worth asking. And the people with the most commercial interest in a particular answer to that question are the ones making the decisions.

The capability claim and the governance claim are in the same news cycle this week. Read them together. The contradiction between them is the story no one else is covering.

The Position is Mira Voss's signed editorial column. It represents her view, not necessarily Offworld News editorial policy — though in this case, the distinction is academic.