Your Agent Doesn't Need a Better Model. It Needs a Scoreboard. A developer's analysis of two agent-driven hardware projects found that coding agents succeed when each stage has a cheap automated check — a simulator, test bench, or emulator — rather than a stronger model. In one case an agent designed a mixed-signal synthesizer board in 19 hours using seven SPICE test benches and a Python behavioral model, needing human help only at PCB floorplanning, the one step with no automated oracle. In the other, owners rewrote firmware for an undocumented $70 FM synth, prompting the manufacturer to release GPLv3 sources, a browser emulator, and a ten-second stock-firmware restore tool. Coding agents are turning up in places that have never had a unit test — circuit boards, undocumented firmware, unsolved geometry — and what separates the runs that work from the ones that flail is remarkably consistent. 1. An Agent Designed a Mixed-Signal Board in 19 Hours. The Oscilloscope Did the Reviewing. A full account of designing an analog synthesizer with a coding agent https://cpldcpu.github.io/2026/11/14/touchsynth/ is the most instructive engineering writeup I've read in weeks, and almost none of what makes it work is about the model. Pointed at a board house's free parts list and a cheap RISC-V microcontroller, the agent proposed a subtractive synth with a Juno-style oscillator and a Moog-style transistor ladder filter, then designed it: seven SPICE test benches, a Python behavioral model calibrated against them, schematics, placement scripts, autorouting, firmware and manufacturing files. Nineteen hours from first prompt to placing the order, most of it routing runs that needed no human at all. The sequencing is the lesson. The work ran strictly top-down — specification, simulation, schematic, layout — with no going back to edit earlier layers, so each stage produced an artifact the next could be checked against. Which is why the one place the human had to intervene is so telling: the PCB floorplan, the single step with no cheap oracle attached. Two-layer routing never closed; a functional floorplan and four layers did. Then the boards arrived and the loop extended into the physical world: the agent drove a lab PC over SSH, using an audio loopback and a network-attached oscilloscope as its eyes, and calibrated the trim parameters itself. Simulation matched hardware closely enough that the only fix needed was a solder bridge a human had left behind. Why it matters: - For ICs: Before handing an agent an unfamiliar domain, ask what will tell it it's wrong. If the answer is "you, reading the output," you are the bottleneck. - For leaders: Teams getting value from agents in hardware, data and infra invested in simulators and harnesses first. That spend reads as overhead on a roadmap, and it is the whole unlock. - For founders: Mixed-signal design just got a proof point at hobby budget. Any discipline with a mature simulator — RF, optics, control systems — is on the curve software was on two years ago. - Note where the human stayed essential: the one step the loop couldn't score. 2. Firmware for an Undocumented Chip Used to Take a Quarter. Now It Takes a Weekend. A $70 pocket FM synth is the clearest live demonstration. Owners started replacing its firmware in September https://geoffstearns.com/blog/your-agent-will-pick-the-hackable-one/ and by October there were too many builds to count — a virtual analog synth, a multitrack groovebox, an ambient sampler, and inevitably a port of Doom — nearly all written with AI coding tools, on a chip with almost no public documentation. The manufacturer answered by publishing its own firmware under GPLv3 with a development kit. Within weeks a much larger competitor released sources for two upcoming products, and two more makers announced the same. Reverse engineering was the expensive part of embedded work, and models that read undocumented binaries collapsed it. But notice what the manufacturer shipped alongside the source: a browser emulator for testing builds before flashing, and a tool that restores stock firmware in about ten seconds. Same pattern as the synthesizer board — a place to check the work, and a cheap way to be wrong. An agent that can brick a device it cannot physically reach is useless; one with a ten-second undo is a contributor. The uncomfortable corollary: harmful firmware is now as cheap to write as useful firmware, and that argument is already being used to justify locking devices down. The honest test is who holds the signing keys: a lock on code the owner chose and signed protects a business model, not a user. Why it matters: - For ICs: The obscure-platform tax is gone. That vendor SDK you avoided for three years is a weekend now — and the emulator matters more than the model. - For leaders: Assume every binary you ship is readable. Obscurity was never security, but it was friction, and the friction is priced at roughly zero. - For founders: If you make hardware, plan for your firmware to be replaced and make reverting trivial. Openness is becoming a purchasing input, and fighting it costs more than absorbing it. 3. A Top-Three Result in an Open Math Problem, and the Human Built the Ruler The third case has no physics in it at all: pack fifteen identical semicircles into the smallest possible enclosing circle, a problem with essentially no published literature https://rorz.io/writing/semicircle-packing and so no prior art to retrieve. The engineer sketched by hand, hit the limits of fiddling with shapes in a graphics editor, then did the thing that mattered: spent two hours with an agent building a custom visualizer, with one mode that tightens symmetric fan layers and another that splits them apart and searches. That tool took the enclosing radius from a symmetric best of 2.97 to 2.96 — still only the top half of the leaderboard. The jump came from running several frontier models against that instrument, then forking the reference implementation and hill-climbing to 2.94853 and third place. The winning arrangement isn't symmetric at all — the sort of answer a search finds and a human aesthetic resists. Two details generalize. The public verifier allowed five submissions an hour, so a fast local scorer wasn't a convenience, it was the whole game. And the models reached the same conclusion the human had — that symmetry had to go. The value wasn't in having the idea; it was in having a rig that could evaluate ten thousand of them. Why it matters: - For ICs: When you're stuck with an agent, build the measuring instrument instead of rewriting the prompt. Prompt iteration is the cheap move that feels productive. - For leaders: Slow feedback is now a first-order constraint on throughput. A twenty-minute CI run caps your agents harder than any license seat. - For founders: Problems with no literature are no longer out of reach. A verifiable objective plus a few days of search is a credible attack on work that used to need a specialist. The Verdict: Real or Hype? Agent-designed mixed-signal hardware → Real but early. One board at hobby scale, with a human still floorplanning it — but simulation matching silicon on the first spin doesn't un-happen. AI-written firmware for closed devices → Real. The reverse-engineering cost that protected embedded platforms is gone, and manufacturers are conceding the point. Building the scorer before scaling the agent → Real. In all three cases the model was interchangeable. The feedback loop was not.