The 99.3% Test for Industrial AI Agents

At QuEra, an Anthropic agent reportedly improved laser stabilization from 58% to 99.3%. The test was narrow, but the number is hard to classify: it is neither merely a model score nor yet proof of industrial autonomy. A model had crossed from recommending an action to changing a machine.

Model developers have lowered the cost of judgment faster than the cost of safely turning judgment into action. Any builder crossing that boundary needs machinery that decides whether an AI agent may act, translates its intent into equipment-specific commands, observes the result, and stops the operation when reality declines to cooperate.

Physical systems expose every architectural shortcut that software can conceal. A malformed document can be restored from version history. A robotic arm colliding with an instrument has a less charming undo function. A control mistake can spoil an experiment, halt production, damage an asset, or injure a person. At this boundary, raw model capability becomes only one input.

Agents matter when they acquire authority, not personality

In 2024, vendors moved from copilots that helped people navigate applications toward agents designed to act on users’ behalf. Buyers could judge a copilot by the quality of an answer. They had to judge an agent by whether it completed a workflow, respected its authority, handled exceptions, and left the system in a recoverable state.

By early 2026, coding agents were completing complex projects with minimal oversight. The frontier was no longer a clever completion but sustained execution: reading a repository, modifying files, running tools, encountering failures, and continuing toward an objective.

When models can produce predictions cheaply, builders must redesign products around decisions, permissions, tools, workflow state, and accountability. The useful question stops being “What does the model know?” and becomes “What can this model cause to happen?”

Anthropic’s Model Hardware Standard, or MHS, makes that shift explicit. The framework is intended to help agents operate physical systems including microscopes, quantum-computing hardware, and robot arms. It does not merely give a model another source of data. It gives the model a route into equipment whose state matters after the token stream ends.

Google DeepMind’s Gemini Robotics 2 approaches the same boundary from another direction, combining multiple models into a system intended to control robots including humanoids. Anthropic emphasizes an interface across equipment; Google DeepMind demonstrates a model system across robot forms. The products differ, but their incentives converge as long-horizon execution improves and integration, authority, and recovery constrain deployment.

Heterogeneous machines make interfaces more valuable than demos

A robot demonstration can prove that one model, one machine, and one carefully assembled environment work together. An industrial control plane has a harder job: it must make many models, instruments, runtimes, and organizational policies interoperable without pretending they are identical.

Laboratories and factories are heterogeneous by default. Equipment comes from different vendors, exposes different command structures, produces different telemetry, and carries different failure modes. Some instruments are new; others are valuable precisely because replacing them would require a capital committee, six signatures, and a minor geological epoch. A clean-sheet universal robot does not solve that installed-base problem.

With a common interface, a laboratory or manufacturer can change models while preserving device integrations. Model providers can orchestrate assets they do not own, while asset owners can avoid making one model vendor permanent merely because its adapter was expensive.

Digital-agent builders have reached the same conclusion. The Linux Foundation’s Agent2Agent project brought together AWS, Google, and Microsoft around a donated protocol specification and software development kits. Microsoft’s Agent Control Specification treats permissions as a separate layer, while OpenAI’s Agents SDK added native sandboxing and a testing harness for long-horizon tasks. MHS extends the same architectural decomposition into physical equipment.

No single specification covers the whole control plane. Builders still need the following functions in combination:

Customers can let a model propose an action without giving it unrestricted credentials to every system involved. They can also replace the model without replacing the identity layer, policy engine, device adapters, logs, or emergency controls. A shared interface therefore gives customers bargaining power as well as technical flexibility.

Physical safety has to be enforced outside the model

A model’s alignment behavior is not a sufficient safety boundary when that model can move a robotic arm, modify a laser, alter a scientific protocol, or reconfigure a compute environment. The model can supply safeguards, but it cannot also be the sole interpreter, permission authority, operator, auditor, and incident responder for its own actions.

Operators must separate deciding, authorizing, executing, observing, and recovering. They need controls at the moment of action, not only in a model card written before deployment.

Microsoft’s control specification treats permissions as an independent layer. World’s AgentKit, built to let websites verify that a real human stands behind purchases initiated by shopping agents, demonstrates the related identity problem. An agent does not acquire legitimate authority merely because it possesses a credential. It needs a chain connecting the action, the agent, the policy, and the person or institution on whose behalf it acts.

Operators should grant an agent the minimum authority required for a bounded task. Systems should validate commands against equipment state and operating limits, observe actions as they occur, and preserve a human override that does not depend on persuading the model to stop. Recovery must mean reaching a safe state, not assuming every physical process can be rolled back like a database transaction.

Enterprise and industrial operators face the same agent authorization and recovery problem, but machinery raises the consequences. Physical systems make vague delegation expensive enough that permissions, identity, auditability, and override paths become core infrastructure.

The first prize is throughput inside constrained environments

Anthropic’s early MHS examples matter less as robotics spectacle than as evidence about the adoption path. The company reported that an agent reduced an imaging experiment at HHMI Janelia Research Campus from weeks to a day and improved laser stabilization at QuEra from 58% to 99.3%.

Researchers can define operating limits and retain supervisory authority in these constrained, high-value environments. The experiment either completes or it does not. The imaging cycle takes weeks or a day. The laser remains stable at a recorded rate. Agents can handle repetitive adjustments, tool sequencing, and error recovery against measurable outputs.

Companies currently use agents primarily to improve efficiency and reduce costs rather than generate top-line growth. Laboratory automation and advanced manufacturing fit that incentive. Their workflows already have expensive equipment, skilled labor, measurable bottlenecks, and a reason to pay for higher utilization.

Anthropic has so far published early tests, not proof of broad production deployment across laboratories or factories. MHS remains a first-phase research preview, and no mature ecosystem of common robotic and laboratory-control interfaces is yet visible. Industrial automation incumbents such as ABB and Siemens are strategically relevant to any broad interface layer, but the launch evidence does not establish their participation.

MHS still has to accommodate installed equipment, satisfy operators, survive procurement, and give vendors a reason to implement it. Early performance shows the prize; it does not establish which standard will capture it.

The control plane can shift power in either direction

A common interface gives model companies access to equipment they do not own, but it also gives customers a route away from the company that created the interface. This is the central strategic tension.

If a specification is genuinely portable, customers can reuse device integrations and substitute model vendors more easily. If one provider’s models, tools, or hosted services work materially better with a nominally open specification, the interface can become a distribution advantage. One provider gains concentration only when customers value its implementation enough to surrender the switching freedom that interoperability creates.

Vendors still disagree about what qualifies as an agent, and that lack of consensus has already created customer confusion. It is difficult to standardize authority for a category whose vendors use the name for everything from a scripted assistant to a persistent autonomous worker. Industrial buyers tend to resolve such ambiguity with narrow scopes, supervision, and procurement paperwork—the traditional predators of exuberant terminology.

Industrial customers therefore do not need every vendor to agree philosophically on agency. They need systems to expose concrete capabilities: read this sensor, adjust within this range, request approval above this threshold, log every command, and stop when this condition appears. Operational contracts can become precise while marketing remains spiritually adventurous.

Anthropic can gain leverage without owning the machines

Frontier labs now reach downward into chips and compute, outward into enterprise workflows, and upward into orchestration standards. They want more control over the stack, but its capital intensity makes complete ownership impractical.

Anthropic has reportedly agreed to pay Nscale $45 billion over six years for roughly 460 megawatts of West Virginia data-center capacity using Nvidia Vera Rubin chips. The agreement remains unconfirmed, so its final form is uncertain. Even as a reported commitment, it illustrates the scale of the compute obligations surrounding frontier models.

Anthropic has also confirmed that it is building an in-house silicon team and pursuing model-and-hardware co-design through a multi-chip approach. Frontier labs want leverage over compute architecture, but direct ownership remains expensive and operationally difficult.

MHS offers a capital-light form of infrastructure leverage. Anthropic does not need to manufacture every microscope, robot arm, laser, or quantum system if it can help define the interface through which agents operate them. The same abstraction that lets customers preserve device integrations lets a model company coordinate external assets without putting those assets on its balance sheet.

By opening a standard, Anthropic can enlarge the market for physical agents while reducing the power of any single model provider, including itself. It gives up some proprietary control to make the addressable system larger.

The 99.3% result changes what performance means

Anthropic’s QuEra result remains an early test, not proof of broad deployment or a winning standard. But the move from 58% to 99.3% changes the performance question. Once an agent adjusts laser hardware, the result cannot be separated from who authorized the adjustment, which limits governed it, what telemetry recorded, and how the system could stop.

A model may choose an action. The industrial system must decide whether that choice is authorized, observable, and recoverable enough to become motion. As model judgment gets cheaper, the scarce industrial asset is the boundary that turns it into accountable work.