Anthropic’s MHS gives us more confidence in physical agents

On August 27, 2026, Anthropic released the first research preview of the Model Hardware Standard, or MHS. It defines standard drivers for microscopes, liquid handlers, robotic arms, laser systems, and other programmable equipment, giving agents a common way to discover, read, and control them.

The questions behind MHS were already familiar to us. Months earlier, seenzus had started working through how an agent should reach devices from different systems, understand which place and room they belong to, retain the state and experience of a physical space, and remain within the authority people have granted it.

Before MHS was announced, seenzus had already built this direction into its spatial model, ingestion protocol, Operation History, Memory, and permission design. Anthropic is now building a standard interface between agents and physical equipment. Two independent paths point in the same direction, giving us more confidence that physical agents deserve long-term investment.

This direction was already part of seenzus

In the months before Anthropic announced MHS, seenzus had already committed to cross-source device access, an editable spatial structure, physical context, device-side outcome confirmation, and action permissions. Product names and implementation details kept changing, but the central direction did not: device control was the entry point, and our destination was an agent that could live with a real place over time.¹

It needs to understand a World, Place, Space, and Device. It reads the scene, remembers how the place operates, revises old experience when reality disagrees, and never treats learning as additional authority.

MHS adds industry evidence for this direction

MHS initially targets scientific research and advanced manufacturing. It is not open source yet, access is limited to a group of partners, and it only works with programmable equipment. Its scope differs from seenzus, but the starting point is close: a model needs an interface that is independent of any one machine before it can operate in the physical world.²

MHS hides machine-specific interfaces behind standard drivers and a small set of read and write primitives. Devices can be discovered automatically. People can add physical information that code alone may not reveal, including the weight of a robotic arm, what an instrument measures, which parameters are adjustable, and the safe operating range. An agent can use those capabilities through MCP, a command line, or an API.

Anthropic also treats the weeks or months often spent connecting laboratory and manufacturing equipment as a cost that shared infrastructure should absorb. MHS aims to reduce some integrations to hours or minutes. Those figures come from the standard’s developer and its early partners, not from a general benchmark. Still, the decision to build a standard moves the agent-to-hardware interface beyond one product team’s internal integration work.

seenzus is designed for homes, shops, studios, and other places people use over time. MHS first serves laboratories and manufacturing. The settings differ, but both treat reliable understanding and control of physical devices as a basic condition for agents. Their independent convergence adds evidence for physical agents as a product direction.

The shared direction appears in concrete design choices

MHS makes devices discoverable, expresses reading and writing through common primitives, and adds physical properties and safety limits. seenzus also keeps device sources behind replaceable ingestion layers, then uses spatial structure and Memory to explain the scene around each device. Both need actions to return real results. Both leave higher-risk operations for people to confirm.

This does not make MHS and seenzus the same kind of system. It shows that agents entering the physical world encounter a similar set of constraints. A device cannot remain trapped behind one vendor’s interface. Sending a command does not mean reality changed. Understanding a device’s capabilities does not grant permission to use them.

The experiments show why physical agents need maintained context

Genentech used MHS to connect a liquid handler, a robotic arm, and a plate reader for a protein concentration assay. The interfaces worked, but physical understanding did not appear automatically. Claude chose the same generic flow rate for ordinary water and a viscous protein solution, which created foam. It then treated the resulting errors as a device problem and retried in the same well, making the foam worse.

Researchers had to explain that this was a physical failure, not a software bug. Claude moved to a clean well, reduced the number of mixing cycles, and retained that context for the rest of the experiment. The team later encoded the lesson as a reusable skill. That sequence needed a device interface, observed results, and maintained experience. It also corresponds to why seenzus puts Operation History, Memory, and spatial context above device ingestion.³

Anthropic’s robotics research found a related effect. The same model controlling the same robot behaved very differently depending on the interface it received. Models mostly failed when asked to issue motor torques directly. Their capabilities expanded when they could use a high-level controller or supervise a pretrained policy. Model weights alone did not determine physical capability. The interface between the model and reality mattered as well.⁴

The QuEra case reached the permission problem. Claude could tune a laser repeatedly through the night, yet an action it considered risky could make it stop and wait for human approval. Too much confirmation leaves an experiment idle. Too little increases physical risk. This corresponds to how seenzus treats authority: learning can improve judgment, but it cannot expand the agent’s permissions on its own.

MHS makes us more confident that physical agents deserve long-term investment

Anthropic’s partners are mostly laboratories and manufacturers, and MHS is still a research preview. The material does not answer whether people in homes or small commercial spaces want an agent that remains present over time. It also cannot prove seenzus product outcomes.

The release offers a different kind of evidence. The seenzus direction had mainly rested on our own product judgment: device counts and protocols would keep multiplying, a single conversation could not carry the accumulated experience of a place, and physical action had to be constrained by observed results and member permission. Through MHS and its partner experiments, Anthropic has turned the most basic premise into infrastructure work on the model side.

MHS shows that reliable access from general models to physical equipment is becoming infrastructure work in its own right. If that infrastructure matures, physical agent products can spend more of their effort on the understanding, feedback, and ongoing relationships above connectivity.

That is its value to seenzus. We had already chosen to build an agent rooted in real places. We now see Anthropic building a hardware path for the same direction. The two paths are independent, and together they give us stronger reasons to keep going.

The direction is unchanged; the confidence is more concrete

The center of the seenzus product has not changed. We will keep device sources replaceable so Home Assistant, MQTT, Matter, or a future MHS integration can enter the same spatial structure. We will not promise support for a specification that has not been published.

The work after connectivity remains the same. The agent needs to know which World, Place, and Space it is acting in. Each operation should settle honestly as success, failure, or unknown. Memory should change when the room contradicts it. Members should continue to decide what the agent may do.

MHS adds evidence for these choices. seenzus still has to provide product evidence. New devices should become part of an understandable place quickly. Long-term context should reduce guessing and repeated confirmation. Users should want to keep what the agent learns and the limited authority they grant it. Those outcomes will determine whether the product works.

More teams building serious agents are beginning to face the same physical world. The seenzus direction was already set. MHS gives us another concrete reason to keep going.

---

Notes and sources

1. seenzus product and version records, reviewed August 27, 2026. The records confirm that the product direction and major design decisions described in the article predate the MHS announcement. This does not imply collaboration, reference, or information exchange between Anthropic and seenzus. Internal paths and version identifiers are omitted from the public article. ↩
2. Anthropic, "Previewing the Model Hardware Standard," August 27, 2026, accessed August 27, 2026. MHS is currently a research preview and is not yet open source. The description of its design, partner cases, and limitations follows this release. ↩
3. Anthropic, "Previewing the Model Hardware Standard," Genentech and QuEra cases, August 27, 2026. Partners reported the foam-related physical failure, expert intervention, and delays caused by human confirmation. A small set of laboratory cases does not predict product outcomes in homes or commercial spaces. ↩
4. Shmuel Berman, Michael Ilie, Jia Deng, and Daniel Freeman, "Claude plays robotics," Anthropic, July 9, 2026, accessed August 27, 2026. The study compares several models, robot bodies, and control interfaces. It supports the claim that interface abstraction affects task performance, not a direct claim about seenzus product outcomes. ↩