Tom Mann, a research engineer at Meta, says he spends his days on software agents that help build the next generation of AI models. He lays this out in a post on X published on 8 November 2026, part year in review and part career guide. It is one employee’s personal account. Meta has announced nothing here, and the post offers no product dates and no results from Meta’s own models.

The idea has a name, recursive self-improvement, or RSI. Mann explains it to family and friends this way: you teach AI agents (programs that act on their own toward a goal) to train the next models and to get better themselves. Each gain is supposed to help the system find the following one.

He sees two routes. One changes the model itself. The other improves what surrounds it, such as the instructions, tools, memory, or code the agent works with. He concedes that “there’s still a lot to figure out.”

The day-to-day sounds plainer than the label. Mann began as a printer repairman in technical support and has no academic background. His team is a mix of job titles (scientists, engineers, data specialists) that mostly add up to the same work. By his estimate, 90 percent of his time goes to writing code and the rest to coming up with ideas and testing them.

His first paper, on what he calls Research Preference Models, asks how to choose which experiments deserve a run when trying them all is too expensive. Mann says the methods he tested beat an unguided baseline, with 95 percent confidence intervals shown in a chart. His text gives no figures for the size of the gain. He adds that many experiments went nowhere and were left out of the paper.

The most useful passage is a mistake. Early in his career he ran an experiment that looked great, and a colleague wanted to know what had shifted the result, and he could not say. He had “vibed too hard,” he writes. Checking afterward, he found a bug that made the results look better than they were. He now prepares answers to the obvious questions before presenting: what changed, why it should help, and what evidence he has.

That story points at a weak spot in the automated version. A bug that flatters a result costs a person one awkward meeting. An agent that selects and runs experiments with little supervision can build on the same flattering number without anyone being embarrassed into checking.

Mann’s advice for people who want this job follows from it. He suggests picking one task, wrapping a small agent harness (the software around a model that feeds it tools and instructions) around it, running a baseline, and changing one thing at a time. Then read the traces, the step-by-step records of what the agent did, and write up what failed. He lists five skills to build: how models are trained, how to orchestrate agents, how to work with coding agents such as Codex and Claude, how benchmarks are designed, and how to read data.

On benchmarks he uses PostTrainBench as the example. Each agent receives a small starting model, a single H100 chip, and a ten-hour budget to make that model better through further training, then gets scored on tests including GPQA, GSM8K, and HealthBench. His instruction is to read the methodology and find out which actions the agent had permission to take before trusting a leaderboard score.

Mann is describing a job that sits close to the frontier of how models get made, and he describes it mostly as careful checking. If agents take over more of the experiment loop, the person who can explain why a number moved becomes the scarce hire, and teams building agent pipelines should budget review time for that person before they budget more compute.

Posted by Tom Mann, a research engineer at Meta, in an article on X on 8 November 2026.