Engineering notes on agents, harnesses, and getting real work out of AI in a domain that doesn't forgive "almost right."
An agent is only as useful as the place it can act. Pasting a schematic into a chat window was never going to work — so we built the substrate: a real PCB design tool, in the browser, git-native, where agents read and write the actual files.
Four words the market uses loosely. Defining them clearly makes the claim legible — and shows where most "AI for X" tools stop.
The workspace is server-side and asynchronous. What that means architecturally — and why "close the tab" is the feature that changes the workflow.
How the harness benchmarks frontier models against real PCB tasks and routes each action to the model that performs best — so you never have to.
Why every agent action lands as a commit, how visual diffs work on schematics and layout, and what rollback across an agent tree looks like.
A design crosses schematic, layout, sourcing, PLM, and fab — each in its own tool. The harness is the connective layer that carries context across the chain.
True sub-agents run their own loops and report back — nesting levels deep. How the parent waits, the child works, and the result flows back up.