ioma8 describes using LLM agents to investigate old software and hardware, drawing on two decades of reverse-engineering experience. The workflow puts saved evidence and human steering around tools that turn machine instructions into something easier to inspect.
- Map the program’s structure and write findings to Markdown as the investigation proceeds. Those notes survive shortened conversations and agent handoffs.
- When an agent gets stuck, record one hypothesis, design an experiment, and save the result in a CSV ledger. Failed explanations become evidence for the next attempt.
- Add breakpoints, logging or controlled execution when inspecting the code alone leaves a question unresolved. Redirect the agent when it keeps following an unproductive trail.
The author links projects involving writing devices, cameras, synthesizers and an old printer driver. These are personal reports, without a controlled comparison of speed or correctness. Recognizing where to steer the agent still depends on the author’s experience.
The 67-comment thread on Hacker News adds practical limits and alternative ways to work.
What the thread adds
- Retr0id — reports repeated guessing on harder targets. They prefer patient tracing when the answer is already in the code, reserving hypothesis testing for genuinely hidden behavior. They also report success having the model write code to run firmware inside an emulator, but warn that “sometimes it tries to take egregious purpose-defeating shortcuts.”
- bri3d — argues that the connection between model and disassembler can be the bottleneck. Their experience favors direct Ghidra scripting or a better tool bridge over assuming the disassembler itself is inadequate.
- crc32 — says a decompilation project improved after simplifying an overengineered workflow. Whether its custom tooling will help another project remains an open question.
- therepanic — values learning through the manual process and finds delegation less satisfying. vbezhenar describes a middle ground: do the investigation manually while the model explains the interface and checks the work.
- saidinesh5 — narrows the agent’s search before handing over a camera-protocol investigation: “instead of going through the full tcpdump logs, i had a 50 line file with only the data it needs, stripped out of all headers.” They also restricted inspection to two libraries and two source files. Their account makes selecting evidence part of the human’s work, alongside saving the agent’s findings.
HN handles are pseudonymous; no per-comment scores are published. Ordering is HN’s own ranking, so this is a slice of the thread, not a consensus.
Daily coverage
Also covered in the November 19 morning digest.