When Retry Made More Bubbles

The liquid handler reported a runtime error.

The agent did what software has trained us to expect from a competent system: it changed a parameter and tried again. The second attempt used the same well. The mixing made more foam. The foam confused the liquid-level sensor, displaced some of the volume the pipette was meant to carry, and distorted the reading that would eventually judge the experiment.

The operation had been retried. The error handler said recovery was underway. Reality had already changed.

This failure appears inside Anthropic's new research preview of the Model Hardware Standard, a proposed common interface for agents operating microscopes, liquid handlers, robotic arms, lasers, and other programmable equipment. An MHS driver gives each device a small shared vocabulary, including commands such as read and write. It can also describe the machine's characteristics, measurements, adjustable settings, and enforced safety limits so an agent can discover and coordinate equipment that previously required bespoke integrations.

The promise is not merely that a model can call a machine. It can watch several machines, alter parameters as conditions change, and assemble what it learns into deterministic code for long-running work. In one case, a laser is adjusted, observed through a camera, adjusted again, and eventually aligned by a script. In another, an agent supervises an experiment across a liquid handler, a robotic arm, and a plate reader.

Then the bubbles arrive.

Genentech researchers were automating a protein assay using liquids with different physical properties. Water tolerated a much faster flow than a viscous protein solution. The agent initially applied generic handling parameters to both, creating bubbles in the thicker sample. When those bubbles produced hardware errors during mixing, its first recovery strategy was to retry with different parameters in the same well. Each retry agitated the liquid further.

A researcher had to supply the missing physical interpretation. The error was not only a failed command. It was foam. Recovery meant moving to a clean well and mixing more gently. Once told, the agent kept that context for the rest of the run, and the team later encoded the lesson into reusable liquid-handling skills.

The agent's interpretation was coherent inside the software. A sensor had raised an error, the operation had not completed, and another attempt might clear the fault. The bubbles were evidence from outside that account. Reality was the source of truth, and the system did not yet know how to read it.

Software culture has made retry feel nearly innocent. A network request times out, a job loses its worker, a test meets a flaky dependency: try the operation again, preferably with a little delay. Well-designed systems still ask whether repeating the action is safe, but the interface often presents failure as if the world has returned to the line before the error.

The physical world keeps the attempt.

It keeps heat, pressure, wear, elapsed time, depleted reagent, a moved arm, a bleached sample, and, occasionally, more bubbles. An action can stay inside every declared limit and still make the next action less safe or less useful. Device-level guardrails matter enormously. MHS can enforce a ceiling on laser power, for example. The current ROS control documentation similarly describes limits for position, velocity, acceleration, and effort.

Those limits answer one question: is this command allowed?

They do not always answer the second: what has the sequence of allowed commands already done to the world?

Robotics systems have long treated feedback and cancellation as part of the interface. In ROS 2, an action is designed for long-running work with intermediate feedback, a result, and a way to cancel. That shape becomes even more important when an agent chooses the next move. The observation after an action cannot be a decorative status update. It is the evidence that decides whether repetition is recovery or escalation.

This is why autonomous laboratories are a richer systems problem than connecting a language model to a robot. NIST describes them as closed feedback loops in which algorithms choose what to make and measure next. Once the loop closes, the quality of recovery depends on what the system accepts as authoritative: not only an error code, but the observed condition that produced it.

The interface should treat reality as the source of truth. A retry record could show which physical state is known to persist, which measurements have gone stale, what material was consumed, and whether a clean starting point still exists. Recovery policies could carry budgets for cumulative exposure, motion, heat, mixing, or elapsed time instead of counting only attempts. Evaluations could include failures where the environment does not politely regenerate between runs.

Human review belongs earlier than the emergency stop. The Genentech example became safer when a researcher's observation overruled the system's interpretation, changed how the error was represented, and then changed the reusable procedure. That is different from leaving a person beside a queue of failed experiments. The expert helped the agent build an account of the work that was answerable to the work itself.

This extends The Tool That Was There To Be Ignored. Restraint is not only knowing which available tool to leave untouched. Sometimes it is knowing that the correct tool, called again, is now the wrong action because the first attempt changed its object.

It also brings the boundary in The Fake Company Had a Real Address closer to the bench. That piece asked whether an agent's story about an environment matched the environment it could actually reach. Hardware adds another demand: the system's internal account of state has to keep up with a world that accumulates consequences even while every command returns a tidy status.

An agent that can touch the world needs more than permission to act and a limit on how far it may move. It needs to recognize when its internal account and the observed world disagree, then let the world win.

The error message said try again. Reality said the first attempt had changed the experiment.