Iluvatar Labs used its AI scientist Marvin to design AgentPad13, an open-source macropad similar to the $230 Codex Micro. The boards cost around $35 each and reached zero routing errors.
Existing tools had plateaued near 95% completion.
Marvin learned PCB routing from research and experiments in 48 hours.
What happened
Iluvatar Labs used its autonomous scientist Marvin to design AgentPad13, an open-source version of the $230 Codex Micro, for around $35 per fully populated board. The project achieved zero unrouted connections and zero design-rule violations within a weekend.
Why it matters
Existing tools like KiCad/FreeRouting and Quilter plateaued at roughly 95% completion, leaving 18-19 of 353 connections unfinished and making unacceptable compromises. Marvin treated routing as a research problem, learning from literature and running iterative experiments, which enabled a complete route within the board's 2-layer and 100mm constraints.
What to watch
The routing skill Marvin developed is planned to be released as an open-source package. AgentPad13's PCB, case, firmware, and build instructions are available on GitHub with a configurator for customization. A known hardware mistake with two underglow LEDs pointing outward has been corrected in the next revision.
Ask the AI about this article →
The AgentPad13 project began as a practical desire for a cheaper, open alternative to the $230 Codex Micro, which was also sold out. But it evolved into a test of whether Marvin, an autonomous scientist, could handle a problem far outside its original scope. The key challenge was PCB routing under strict constraints: a 2-layer board (to avoid a 50% price increase for 4-layer) and a 100mm max dimension.
Marvin's approach was to stop treating routers as an oracle and instead synthesize knowledge from literature, datasheets, and working keyboard layouts into a runbook. It then ran iterative experiments, measuring progress against the complete design rather than just the number of unrouted connections. This prevented it from making the same mistakes as previous tools, like breaking USB connectivity to reduce unrouted totals. Key learnings included routing power early and wide, reserving corridors for critical signals, and spreading out the MCU cluster to allow traces to escape.
The final board passed all manufacturing checks and the firmware was simulated before hardware arrived. While the boards work, Marvin made one visible hardware mistake with two underglow LEDs oriented incorrectly, a footprint error that has been fixed. The project also plans to release the routing skill as an open-source package, suggesting the workflow could be reused for other complex design problems.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
Ask AI anything about this article. Q&As are published on this page for other readers too.
The Linux Foundation announced the contribution of TRACE (Trust, Runtime Attestation and Compliance Evidence)…
A large-scale study of 17,022 skills from SkillsMP, the largest open-source skill marketplace, found 520 affec…
AWS announced a redesigned SageMaker Python SDK v3 that replaces framework-specific estimator classes with a s…
Caltech Professor Anima Anandkumar developed FourCastNet, an open-source AI weather model that rivals traditio…
Bedrock Robotics Inc. has deployed fully autonomous, retrofitted heavy excavators onto active infrastructure p…
Goldman Sachs Global Institute co-head Jared Cohen argues that the world will be redesigned for machine intell…
