Older entries
Launches, releases, funding and papers about agents doing hardware work.
Newest first, as we log them.
3 mo ago
Researchers had coding agents write robot behavior trees from a contract served by a robot-side MCP server. The contract lists the robot's skills, allowed operators and optional templates, and a gate validates each tree before it runs. The paper reports tests of Sonnet 4.6 and Gemma4:31b on 110 simulated tasks and 14 on a Husarion Panther running Nav2.
3 mo ago
UST announced an alliance with Anthropic that puts Claude into the engineering platforms UST runs for clients. In UST-iDEC, its hardware and silicon validation platform, Claude Code reads chip pinouts and schematics, then writes and runs regression test scripts that engineers used to write by hand. UST says the platform already cuts validation cycles by 50 to 70%.
3 mo ago
HARDWARIO published a repository on 30 June that holds its TOWER IoT kit's protocol, STM32L083 firmware, CLI, flasher and hardware-in-the-loop bench. It began as a Claude Code control plane over submodules and is now a Rust monorepo with agent instructions. The README says bench runs flash a physical Core Module and measure power with a PPK2.
4 mo ago
Researchers put an LLM agent in charge of a trapped-ion experiment's ARTIQ control code through an MCP server. Every tool call needs a token bound to its contents, issued by a human or after a simulated run stays within device limits. The paper reports that the agent built a calibration stack and, with operator guidance, closed a magnetic-field stabilization loop.
4 mo ago
Anthropic repeated its robot-dog experiment with Claude Opus 4.7 in Claude Code, a researcher approving each command. On every task a 2025 human team had completed, including connecting to the camera and lidar, Claude was at least ten times faster. Anthropic says its attempts to knock the beach ball back were poorly controlled, and it never finished the autonomous fetch.
4 mo ago
Zhihan Zhang and colleagues, mostly at the University of Washington, published Embedded Arena, a benchmark that scores agents' model and firmware edits on microcontrollers. Boards include a MAX78000, an ESP32-S3 and an STM32N6. The paper reports Claude Opus 4.7 and Gemini 3.1 Pro failing every deployment without hardware feedback; with it, the first success came within 3 iterations.
4 mo ago
H2Loop launched Hydron on Hacker News, an embedded coding agent for VS Code and the terminal that answers from pre-indexed datasheets. The launch post lists more than 580 indexed platforms and peripherals, including STM32, ESP32 and RP2040 parts. An agentic serial monitor shipped at launch; GDB integration and a log reader were promised within 2 weeks.
5 mo ago
Memfault released an MCP server with 12 read-only tools covering devices, issues, metrics and software versions. In the launch post, Claude Code finds the day's worst new crash, matches its stack trace to the checked-out firmware and drafts a fix. A GitHub MCP server opens the pull request. Memfault says the tools cannot change devices or trigger OTA updates.
5 mo ago
Nordic Semiconductor released 2 MCP servers that connect coding assistants such as Claude Code to its SDK and cloud. The Nordic MCP server adds documentation search, an nrfutil reference, a UART monitor script and SDK workflow commands. The nRF Cloud server reads fleet data such as logs and coredumps, without write access. The post shows no hardware runs.
5 mo ago
Chunhui Li and Wei Fan describe RFIA, an RF instrument agent whose LLM plans while a deterministic runtime operates the instrument. A knowledge base supplies verified skills, workflow templates and SCPI command lookup. On a commercial vector network analyzer, the paper reports all 16 benchmark tasks handled with both MiniMax-M2.7 and Qwen3.6-27B, including 1 expected safety rejection.
5 mo ago
Penglin Dai and co-authors describe AutoMCU, an LLM multi-agent system that designs neural networks to fit a microcontroller's RAM and flash limits. Vendor toolchain checks reject candidates that would not fit before any training. The paper reports competitive CIFAR accuracy after 1 to 2 hours of customization and tests on several STM32 parts; no code is linked.
5 mo ago
Craig Schindler, CTO of Assemble Labs, describes how the company has AI agents write experiments for embedded devices and interpret sensor and power data. The 10-minute interview was recorded at NVIDIA GTC 2026. His main case study is a wearable health sensor, including how power regressions are found and fixed. The video shows no bench footage of the agents.
5 mo ago
Espressif published a recap on 14 May of ESP32 projects from Anthropic's Code with Claude event, where each attendee got an ESP32-S3 M5Stack Cardputer. Attendees described a program in plain language, and Claude Code wrote and flashed it. The post also covers Claude Desktop Buddy, a BLE reference design on the M5StickC Plus, and Espressif's own ESP Desktop Buddy library.
5 mo ago
Root Access, a US maker of electronics for drones and robots, opened its Seb coding agent to the public on 8 May. In a Show HN post, founding engineer Vignesh K, a former TI applications engineer, says Seb reads schematics, datasheets and BOMs, mixes deterministic algorithms with LLM calls and connects to debuggers and logs. We saw no demo.
5 mo ago
Keysight has released version 0.1 of its MCP Server for Instrument Control as a Windows executable. The release notes list 18 tools that discover VISA instruments, resolve SCPI commands from each instrument's definition file and run them. An engineer approves a plain-language preview of every command before it reaches hardware. Sessions are logged, and a simulation mode needs no instruments.
7 mo ago
AllSpice published a paper on DRCY, a multi-agent LLM system that checks schematic connections pin by pin against component datasheets. It runs as a CI/CD action when an engineer submits a design review on AllSpice Hub and posts inline comments. A consensus agent reconciles repeated runs. The authors say it runs in production; the paper gives no accuracy figures.
14 mo ago
Embedder launched its firmware coding agent on Hacker News on 15 August 2025 with a demo adding a magnetometer to a Pebble 2 Duo. Founders Bob Wei and Ethan Gibbs said it reads uploaded datasheets and schematics, checks output over a serial console and steps through code with GDB. SEGGER announced J-Link and J-Trace probe access in September 2026.
15 mo ago
BootLoop, a Y Combinator Summer 2025 startup founded by Noah Pacik-Nelson and Chris Markus, is building an AI agent that writes and tests firmware. The company says the agent reads datasheets and schematics, then validates code on customer boards with oscilloscopes, logic analyzers and debug probes. In pilots, its own engineers run the agent and a person reviews every line.
