Claude designed three devices: the boards worked without schematic changes

The board is ready, the firmware is written and the enclosure is printed—but the device cannot be assembled: a switch hits the enclosure wall. In hardware development, a mistake like this can send a team into another design cycle. In an Anthropic experiment, Claude spotted a similar problem in the digital model and proposed inserting the board at an angle of about 25 degrees. It is a good example of why electronics developers might need AI: to coordinate decisions about schematics, enclosures and assembly before ordering parts.

According to Anthropic, the company worked with board manufacturer Diode Computers to test Claude Fable 5.2 on three devices: a USB speakerphone, a home voice assistant hub and a pocket virtual pet. Claude prepared the electronics, enclosures and firmware. People fabricated and assembled the prototypes under its guidance. The authors say no changes to the electrical schematics were needed.

Why the final traces make all the difference

An almost finished board design is not enough. If a few connections cannot be routed with the required clearances, the device remains a design on a screen. In earlier experiments, Diode received board designs from models that were 95% complete but unusable.

In the new test, Claude revisited earlier decisions. The speakerphone, for example, has a touch-sensitive ring. Moving between PCB layers requires a via, which in this area could pass through the sensing electrode. The model reassigned signals to chip pins so that the traces beneath the ring could run on a single layer without crossing.

The practical implication is clear: an agent needs permission to revisit component placement and pin assignments. Otherwise, it may spend a long time trying to finish a routing layout that its own decisions made difficult.

Anthropic also reports a numerical result. In Diode's internal PCB Gym test, which requires placing components and restoring connections, Fable 5.2 scored 69%, compared with 25% for Fable 5. That is an increase of 44 percentage points. But this score should not be read as the probability of producing a working device: the test evaluates specific PCB design tasks.

When an agent has to write its own tool

For the voice assistant hub, Claude initially used the KiCad-based tools provided by Diode. The built-in routing tool laid a trace between selected points, but handling long connections from a 100-pin connector required too many individual actions.

The model then wrote its own algorithm: it divided the board into a grid and searched for an unobstructed path for each connection. According to Anthropic's account, in about half an hour it produced a tool capable of routing a single long trace in roughly one second. When Claude noticed a sensitive audio synchronization line running too close to Ethernet, it modified the algorithm to increase the clearance around such signals.

This is a useful detail for teams building engineering agents. Access to CAD controls may not be enough. The ability to write a small helper script makes it possible to automate a repetitive part of the work. The script's output must still pass design checks.

The model also accounted for the relationship between electronics and firmware: on the speakerphone, LED brightness decreased as volume increased, keeping the device within its specified current limit.

A working prototype still has to become a product

Claude checked manufacturing constraints and component availability. For example, it corrected six overly narrow sections of solder mask on the hub's board by enlarging the openings in the coating. According to Anthropic, Diode manufactured all three boards directly from the supplied files, without further rounds of change approval.

The evidence is still limited, however: three selected devices and a report from the experiment's participants. There are no data on long-term operation or successful commercial certification. The report also does not provide enough information to assess the full cost of this development process or promise a specific reduction in development time.

A sensible starting point for a hardware team is a limited experiment: have an agent prepare a small prototype, retain the revision history and check the schematic, routing, enclosure and firmware using standard engineering methods. The main result of this work is a demonstration of a coherent process: the model brought several interdependent designs through to physical assembly. How reliably this can be repeated with other devices still needs to be tested.