Hardware Agent Runtime connects external coding agents to embedded devices through observable, verifiable hardware-in-the-loop workflows.

Software assumptions meet physical devices
An AI coding agent can change a file in milliseconds. A physical board has a cable, a bootloader, a serial port, and the possibility that a wrong action changes more than the code.
Hardware Agent Runtime is a local-first runtime for that boundary. It lets an external agent compile, flash, observe, experiment with, and verify real embedded hardware while keeping the workflow explicit.
A workflow built around evidence
The first target is Arduino through Arduino CLI. The runtime provides a hardware-in-the-loop flow that covers discovery, compilation, flashing, serial diagnostics, and verification reports. The interface is exposed to AI agents through an MCP server over STDIO with input and output validation.
The important design decision is that observation and inference stay separate. A report can say what the device actually emitted, then separately record what an agent thinks that signal means. Inference should never overwrite the observation that made it possible.
Human pauses are part of the system
Some actions should not become invisible automation. Experiments can persist a human-action pause and a safety analysis, so a person remains able to inspect a physical step before it happens.
This also changes what “success” means. A build passing is not enough. The runtime needs attributable evidence that the intended board was discovered, the program was flashed, the device was observed, and the result matched the experiment.
Designing for more than one board
The current first target is Arduino, with adapter contracts that preserve a path toward ESP-IDF, Zephyr, RP2040, STM32, and MicroPython. The adapter boundary is less about predicting every future board and more about keeping the agent-facing workflow stable while transport and toolchains change underneath it.
The runtime has been validated end-to-end on real ESP32 Dev Module hardware. That physical check is the dividing line between a plausible abstraction and a useful one.
What the boundary teaches
Bringing agents into the physical world requires slower, more attributable actions than working inside a repository. Hardware Agent Runtime is an attempt to make that caution programmable: observe first, act deliberately, and treat the device as the source of truth.