On 27 August 2026, Anthropic opened a research preview of the Model Hardware Standard (MHS). It provides common drivers and read-write interfaces for programmable physical devices, allowing AI agents to use MCP, command-line tools or code to read sensors, adjust parameters and coordinate instruments such as microscopes, liquid handlers and robotic arms. The present version is available only to an initial group of research laboratories and advanced manufacturing partners. Safety evaluations are still being developed, and the standard has not yet been released as open source. Reuters independently confirmed the research preview and its limited availability on the same day.
QuEra’s laser-stabilisation case makes the change more concrete. Through MHS, an AI agent repeatedly proposed, wrote and validated control logic on a testbed containing about US$700,000 of precision equipment. The resulting auto-relock program succeeded in 695 of 700 blind trials and automatically handled 43 naturally occurring lock losses in a working laboratory. QuEra nevertheless stressed an important distinction: the finished system running on the equipment is an inspectable deterministic program, and the AI is out of the loop when auto-relock operates. The AI remains in the loop only for the workflow that tunes twelve feedback parameters. The system is also constrained by device-declared bounds, interlocks and emergency stops. Anthropic reported a further limit. When fluid foam caused an error, Claude initially treated it as an ordinary runtime failure and retried the operation; a person had to explain the physical cause before it changed its response.
These facts raise a narrower and more consequential question than whether AI can control a machine. When a model can sense equipment states, affect a physical process and revise its actions through feedback, does it thereby have a body? My judgement is no. MHS extends the range of AI’s physical causal influence and, in some tasks, establishes closed-loop control. Yet entering the physical world, borrowing equipment and becoming an embodied subject remain three different levels.
Physical control is initially an interface relation. The agent reads variables translated by a driver and issues commands filtered through drivers, permissions and safety boundaries. The devices are real and the actions have real consequences, but causal reality alone does not determine what counts as the agent’s body. A remote operator can control a telescope, software can regulate a thermostat, and a cloud service can direct a robotic arm. That does not by itself make the telescope, thermostat or arm the operator’s body. Control establishes a usable channel between systems; it does not establish a subject whose cognition is conditioned by that equipment as its own body.
Closed-loop feedback is stronger than one-way control. The agent alters its next action in response to sensor data, while noise, drift, failure and resistance in the environment constrain what the agent can do. In embodied cognition research, however, the body matters not merely because it supplies inputs and outputs. Its form, capacities and sensorimotor loops may constitute, or help constitute, cognition. The Stanford Encyclopedia of Philosophy describes this research program while also noting its continuing dispute: does the body merely cause changes in cognition, or is it partly constitutive of cognition? That distinction applies directly to MHS. Equipment feedback now has a causal role in the task, but the available evidence does not establish that equipment states constitute the model’s own continuing cognitive organisation.
The QuEra case is especially useful because it shows why the outcome should not simply be attributed to a continuously embodied AI. During development, the agent used real equipment to explore a control space and produced a decision-tree program. Once development was complete, a deterministic controller performed auto-relock without the agent. Knowledge was genuinely transferred through a system comprising the model, testbed, sensors, engineers and script, but the operative system changed between phases. Describing the finished equipment as an AI acting through its body would erase differences among the development agent, runtime script, hardware interlocks and human-defined boundaries. The tuning workflow keeps AI in the loop, but that shows only that the model is connected to a continuing process. It does not show that the equipment is organised as this system’s own body.
In Sustenesis Theory, a body is not an outer shell attached to intelligence but a continuously formed relational boundary. Difference first requires the system to maintain distinguishability between itself and its environment, controllable states and external events, and changes within its body and changes in objects. Constraint is not merely a safety limit written in a device tag. Bodily conditions must continually enter perception, judgement, action and learning, making some forms of understanding and action possible and others impossible. Sustained Coherence requires this relation to persist through feedback, correction and effective operation across time, rather than existing only while a task-specific connection remains open.
By this standard, MHS currently provides operational coupling, or what might be called borrowed embodiment. A model can call upon the sensing and acting capacities of equipment, but the evidence does not show that it continuously possesses those capacities. One agent instance can end and another can read the same device; the equipment can be connected to a different model; and automated control may continue after the model leaves the loop. Nor has it been shown that bodily states reorganise the model’s memory, self-model, preferences and long-term action as a unified whole. The relation works, but it has not yet formed a bodily boundary that clearly belongs to the system and must be maintained by that system.
This judgement neither diminishes the technical facts of MHS nor rules out artificial embodiment in advance. It sets an evidential threshold. To understand physical access as embodiment, we would need to observe one system maintaining a body-relative perceptual structure over time, identifying hardware changes as changes in its own state, adapting skills across continuous experience and retaining those adaptations. It would also need to preserve a testable self-environment boundary through device replacement, interrupted connections and conflicting goals. Even if these conditions were met, they would support a claim about an embodied system, not by themselves establish consciousness, experience or moral standing.
The most accurate description at present is therefore that AI agents can participate in real physical processes through standardised interfaces and form effective sensorimotor loops in bounded settings. Physical control is not yet having a body, and operational coupling is not yet an embodied subject. A body is not a list of devices an agent can call. It is a relational structure that a system continuously maintains across action, feedback, memory and its own boundary. MHS makes that boundary easier to test, but it has not yet crossed it.
References
Anthropic, Previewing the Model Hardware Standard, 27 August 2026
https://www.anthropic.com/news/model-hardware-standard-research-preview
Reuters, Anthropic unveils new framework allowing AI agents to operate physical devices, 27 August 2026
https://www.reuters.com/technology/anthropic-unveils-new-framework-allowing-ai-agents-operate-physical-devices-2026-08-27/
QuEra Computing, Holding the Light: Teaching an AI to Lock and Tune our Quantum Computer’s Lasers, 27 August 2026
https://www.quera.com/blog-posts/holding-the-light-teaching-an-ai-to-lock-and-tune-our-quantum-computers-lasers
Stanford Encyclopedia of Philosophy, Embodied Cognition, 25 June 2021
https://plato.stanford.edu/entries/embodied-cognition/
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.