Speedrun AI Labs Releases Reverse Engineer Anything, an Open-Source Skill With a Recursive Self-Learning Loop

Writes down what 100% means before an AI agent builds anything, then loops build, review, improve until every must-have passes, with a human signing off done.

SYDNEY, NEW SOUTH WHALES, AUSTRALIA, November 10, 2026 /EINPresswire.com/ -- Speedrun AI Labs, the Sydney AI lab that builds cybersecurity agents and compliance agents, has released Reverse Engineer Anything with a Self-Learning Loop, a free, open-source skill for AI agents under the MIT licence. The skill is built for teams that examine a product they are authorised to rebuild, build their own version, and keep improving it against the original until it is complete.

Ask an AI agent to rebuild a shipped product and the result is often 80% there: it looks right and the obvious paths work, then it stops, because nobody wrote down what 100% means and nothing measures the build against the original. The skill treats that last 20% as a method.

Before any building starts, the skill turns its examination of the original into a parity checklist: one testable behaviour per item, each with a concrete input, an expected output and the check that will verify it, tiered as must-have, should-have or out of scope. The operator signs the checklist, and each run is scored as must-haves passing over must-haves total. The build follows clean-room discipline, so whoever writes the code does not have decompiled text open, and each review runs the same input through the original and the build wherever the original can be run, with a pass, fail or not-testable result and the file or command that proves each one.

The skill is done only when every must-have passes in two consecutive reviews, and it never declares 100% on an untested item. It stops and asks the operator when the score stalls for two iterations, when the budget is spent, or when closing a gap would need protected material or a vendor's backend.

After each iteration, the self-learning loop saves checklist patterns, review techniques and fix recipes back into the skill, staged for the operator's approval, so the next project starts further along. It never saves secrets, personal data or any text from the target product, and it never activates a rule it wrote.

The skill is not built for cloning a third party's product, defeating access controls or lifting protected assets. Before any tool runs, it asks for the authority record and the reuse boundary, and it requires an approver independent of the person making the request.

The release combines two earlier Speedrun AI Labs skills, Reverse Engineer Anything and Hermes Agent as a Skill File, and adds a completion loop with a written definition of done. It installs in Claude Code and OpenClaw and works with any agent that has a terminal, file access and a writable skills directory, and it joins the company's free Skill Library, which now holds seven skills.

"Anyone who has asked an AI agent to rebuild software has hit the same wall. It gets 80% of the way and stops, because nobody wrote down what done looks like. This skill writes it down first, then keeps working until every must-have passes. It is built for teams rebuilding software they own or are authorised to rebuild, with a human signing off the definition of done and every lesson the skill keeps."
Adnan Tanveer, founder of Speedrun AI Labs

ABOUT SPEEDRUN AI LABS
Speedrun AI Labs is a Sydney AI lab that builds cybersecurity agents and compliance agents. It ships two: Anton, a cybersecurity agent, and Joseph, a compliance agent for financial services businesses. A human stays in the loop with both. It also publishes a free, open-source Skill Library for AI agents.
www.speedrunlab.ai

MEDIA CONTACT
PR Team for Speedrun AI Labs
info@speedrunlab.ai

DISCLAIMER
Reverse Engineer Anything with a Self-Learning Loop is released under the MIT licence and provided as is. It is intended only for products the user is authorised to examine and rebuild; users are responsible for confirming that authority and complying with applicable laws and licence terms.

The skill is inspired by REA (Reverse Engineer Anything) by @morluto; it is independent of and not affiliated with that project. Hermes Agent is a project of Nous Research. Hermes Agent as a Skill File is a derivative released with attribution; Speedrun AI Labs is not affiliated with Nous Research. Claude Code and OpenClaw are products of their respective owners, which are not affiliated with Speedrun AI Labs and have not sponsored, endorsed or approved this release.

This release contains statements about Speedrun AI Labs' plans. They reflect current intentions and are not guarantees of future products or results.

Adnan Tanveer (Addy)
Speedrun AI Labs