Allspice reports that its hardware design agent beats Claude Code and Codex when given schematic-editing tools

Allspice built DRCY, an agent that reviews hardware schematics and, in its new version, also modifies and creates them. The first version was a fixed sequence of specialized models; the new one uses a harness written in Rust that runs server-side in sandboxed containers. Tirmazi says coding-agent habits did not carry over, because a schematic is not a text file an agent can edit and run. DRCY's harness instead gives the agent a Hardware Mutations Interface, a set of well-scoped tools that change a schematic without the agent parsing its underlying ECAD format.

Allspice says this lets the agent create correct designs at a higher success rate than off-the-shelf harnesses using the same underlying model. The comparison used Claude Code 2.1.322 with Claude Opus 5.5 and Codex CLI 0.167.0 with GPT 6.1 Sol, on a randomly sampled subset of Allspice's own human-verified tasks that the company says was held out from training and tuning. The off-the-shelf runs did not have DRCY's mutation or platform tools. In one example, drawing a voltage divider from two 10 kΩ resistors, the post says Claude Code placed it past the sheet's right border and Codex drew it without resistors.

The success rates are in charts, and we did not read the values. The evals are internal and the author is the vendor. Because the comparison removes DRCY's tools from the others, it measures the tool interface and harness together, not the model.

Why it matters: For anyone building an agent over something that is not plain text, the post's claim is that a small intent-level tool layer beats handing a general harness the raw files. That claim rests on one vendor's internal eval, so treat it as a design to test against your own tasks.