CopperPilot's PCB agent needs to preserve the reasons behind the circuit
CopperPilot has published an internal KiCad benchmark for Epsilon, its AI agent for electronics design. Its published KiCad benchmark offers a useful way to examine what the agent can do, because it follows individual design tasks rather than judging the appearance of a finished board.
The internal test covered 341 tasks across 49 designs, ranging up to an eight-layer Jetson Nano carrier with 370 components. Schematic capture, placement and routing were graded through KiCad's electrical and design rule checks, netlist export and scripted checks on the files. Using the same underlying language model, Epsilon scored 83.2 overall against 54.7 for a general-purpose agent, suggesting that the software around the model made a substantial difference within this test.
The interpretation depends on what those checks can establish. A design rule check can detect a route that breaks a specified clearance, but it relies on somebody having established the right clearance in the first place. An analogue node that needs separation from a switching path may pass every configured check if that requirement never reached the layout data. The resulting score measures success against the benchmark's checks, leaving electrical behaviour to be established through the relevant analysis and validation.
Keeping the design intent through a change
CopperPilot says Epsilon can consult datasheets and reference designs while retaining design intent through longer sessions. Retaining that intent becomes especially important when a decision changes halfway through the work, because an apparently local substitution can affect a circuit well beyond the replaced component. A regulator provides a straightforward example. Another device might fit the footprint and supply the nominal output voltage but require a different capacitor network or behave differently during startup. An agent that preserves the connections could still leave an unsuitable circuit; one that carries the design requirements through the substitution would revisit the supporting components and identify what needs to be checked again.
There is value in automating that repeated work. Finding the dependent choices, updating them consistently and leaving a usable record could save time even when an engineer retains responsibility for judging the result. The benchmark's use of native EDA checks helps explain part of the workflow, although CopperPilot has reported the results itself and the scores do not establish how many completed assemblies have passed electrical testing.
For Epsilon, preserving the reasons behind a component choice is therefore a more demanding task than preserving the netlist. A circuit may remain connected correctly after a substitution while the assumptions that made it work have changed.
Sources
Image: Visual inspection of an assembled circuit board. Illustrative photograph; not the CopperPilot software or its output. Photo: Encik Tekateki / Wikimedia Commons. Image source. CC BY-SA 4.0.
