Notes

REA is an open-source project by morluto that connects AI agents to reverse-engineering tools via CLI and MCP, assisting in the investigation of application behavior, source code, and native programs. I appreciate how this project transforms reverse engineering into a dialogue with a programming agent: you can ask how a specific feature functions, follow the trail of evidence through the code or runtime behavior, and ultimately decide how to implement it yourself. REA’s emphasis on evidence makes its conclusions significantly more reliable. For individual developers, this lowers the barrier to studying legacy software or addressing compatibility issues. If you simply want to add a small, convenient feature to an existing application, reverse engineering not only makes it easier to achieve but can also result in a more flexible implementation. It suggests a future where software engineers might frequently begin developing new features by "reading" existing software, rather than starting from a blank file.

A major value of REA lies in its ability to integrate standard reverse-engineering tools (such as IDA, Ghidra, or Hopper) into an AI agent. This provides those with no prior experience in reverse engineering a "camera" to capture exactly what happens inside an application when a specific button is pressed. Instead of guessing how a favorite feature was built, an agent can now provide concrete evidence, walking you through the implementation step by step.

However, as reverse engineering becomes more accessible, products created by individual developers may become increasingly easy to replicate. Since independent developers lack the "moats" enjoyed by big tech companies - and have fewer means to protect their work - they may increasingly need to rely on robust legal frameworks. In an era where technical barriers and the difficulty of replication are steadily declining, I believe that the quality of the core idea will become more important than ever. For instance, consider a popular mini-program that gained traction recently: you post a difficult decision you’re grappling with, and it’s sent out like a "message in a bottle" to a random stranger. That stranger then offers advice and their reasoning on whether or not you should proceed. There is something uniquely persuasive about receiving counsel from a complete stranger, and it often leads to surprisingly interesting perspectives.

In September 2026, Meta, Manus, and OpenAI launched Muse, Cue, and dots, respectively, all designed to enable personal agents to handle ongoing daily and professional tasks. These products suggest a shift in the role of agents: moving from merely "helping me answer" to "taking over and executing the work." Muse is geared toward everyday life, such as shopping or travel, but incorporates independent security reviews, drawing a clear line between "handling a task on your behalf" and "authorizing the actual execution." Cue equips personal agents with access to phone, email, and computer systems, while dots functions more like an advanced office tool, also emphasizing continuous, long-term work. I am particularly drawn to its philosophy of "giving it a responsibility to own." In my experience using various coding agents, even with a well-prepared specification or plan, they often get stuck on minor details or grow increasingly "lazy" over time, requiring me to constantly nudge them and steer them back to the original roadmap. In contrast, dots is said to be capable of proactively determining the next logical steps and executing them, leaving only key decision-making to the user. I’m eager to test this by handing over some "vibe coding" projects to it. What I’m most curious to observe with these three products is whether, in the long run, people will truly entrust their to-do lists to them, and how frequently they will need to step back in to check on the progress.

I followed Anew Labs’ AnewDDE technical report. Anew Labs spun out of ByteDance’s internal drug discovery team, and ByteDance retained a majority stake after the financing. What caught my attention is how AnewDDE links protein structure, affinity evaluation, molecular design, and R&D reasoning in an experimental loop. In its VHH nanobody affinity-maturation round, 16 of 100 candidates measured in the single-digit nM range. I find that more interesting than a model leaderboard; next, I want to see whether similar hits recur across targets and translate into function. Beyond AnewDDE, the suite has gradually formed a comprehensive pipeline. Anew Labs does not merely provide an AI platform; rather, it aims to establish its own end-to-end drug discovery pipeline.