Simantic, a Y Combinator F26 company, is building firmware simulation infrastructure aimed squarely at coding agents. The pitch is narrow and concrete: give the agent a virtual development board it can program and run through a single API call, so it stops reasoning about registers and starts observing behavior.
The problem with asking an agent to write firmware
Embedded code is hostile to the usual agent loop. A web agent can run the test suite and read the failure. A firmware agent writes a register write, and then what? The feedback loop disappears at exactly the point where correctness matters, which is why embedded work has stayed resistant to agents that do well everywhere else.
Static review does not close that gap. A diff can look correct and still leave a pin in the wrong state, because the error lives in timing and sequencing rather than in syntax.
What the launch actually demonstrates
The demo targets an STM32F072RB and reports a concrete finding: a pin labeled PA5 that never goes low. That is the right shape of output. It is not a style complaint or a plausible-sounding summary; it is an observed signal that contradicts the intent of the code.
The company also describes a pull-request workflow where the agent builds the firmware, generates a scenario from the diff, runs the simulation, and posts a single review comment. In CI, the same mechanism validates each commit against a simulated MCU.
Why the earlier approach was abandoned
The team says an early MVP wrapped Renode, an existing open-source simulation framework, and that they moved away from it because merge velocity and modeling depth were both constrained. That detail is worth noting: it is the difference between a wrapper and an engine, and it explains why the current product can expose a board-level API rather than a generic emulator session.
Where the limits are
Simulation is only as good as its models. Peripherals that are not modeled will not fail in simulation and will fail on the bench. Timing fidelity is an approximation, and analog behavior is largely out of scope. The vendor’s cost comparison, which cites a satellite development board above 100,000 dollars being validated at roughly one hundredth of that cost, is a vendor claim rather than an audited result.
The realistic reading is that this shifts the cheap part of validation earlier. It does not replace bring-up.
What this signals for agent tooling generally
The interesting move is not the specific board. It is the pattern: agents get better when someone builds the execution environment that turns their output into an observable result. Code agents needed a test runner. Firmware agents need a board. The companies that build those environments are, in effect, building the feedback loops that make agents useful in a new domain.
If you want to see how this fits alongside desktop agents that drive real machines, the guides index collects the operational write-ups, and the posts section tracks the tooling news as it lands. To run an agent on your own Windows machine, start at the download page, and read is an AI agent safe for my PC before granting file access.