Jordan fed his raw 23andMe file and blood work to Claude Code for a local health dashboard. Plus Sam's Codex-built AI trainer and the rise of multi-agent AI.
Yes, Claude Code can read a raw 23andMe file and turn it into a personal health dashboard. On Episode 30 of Built This Week, co-host Jordan Metzner showed one he built from his 15 MB DNA export plus his blood test results and medications, running only on his own computer. Sam Nadler also showed an AI personal trainer he built with the new Codex app, and we talked about why multi-agent AI is the biggest shift in how people use LLMs.
Jordan had done 23andMe a few years earlier. When the company looked like it might go bankrupt and people worried about their data being sold, he downloaded his raw file: a single text file of about 15 MB that he calls "pretty much unreadable." It sat in his Google Drive until he went looking for a new project and decided to give Claude Code access to his DNA.
The first pass looked like what you would see inside 23andMe itself: APOE status, diabetes risk, lactose intolerance, stress responses and aging markers, then deeper detail across his chromosomes. Not every result matched real life. It flagged high lactose intolerance, which Jordan says he does not notice in himself. He then added his blood test results and the drugs he takes, so the system could tie his personal health data to his DNA.
One insight became a family project. A cholesterol medication had a big effect on Jordan's LDL, and Claude suggested he talk to his blood siblings. His sister had never heard of the medication and asked what she should tell her doctor, so he had Claude Code write her a report to bring to that appointment.
So I actually had Claude Code build an entire report for my sister based on my DNA and my health information so she could share it with her doctor, and so that they could analyze her health against what might be hereditary or kind of inherited behaviors.
— Jordan Metzner, Episode 30
Jordan's point was not that a dashboard replaces a doctor. The decision stays between his sister and her physician. What interested him is that siblings rarely share health information, and a report like this could become a new way for families to get a deeper look at how DNA affects their health. You can see the build on our personal DNA and health insights dashboard page.
Watch Jordan's DNA dashboard in Episode 30 (starts at 1:38) · Episode 30 notes and transcript
Sam asked whether the dashboard sits on a backend database. Jordan thinks it uses SQLite, and it runs locally on his computer only. He has no interest in putting it on the internet. He opens it when he needs it, and he plans to show it to his doctor at his next visit, which was a few weeks away at the time of recording.
The plan is to keep it current. Every time he gets blood work back, he adds it, so the system holds everything in one place. He also hopes newer models will surface insights that were not easy to get before.
Getting the data in is easier than it sounds. Jordan says that if your doctor uses something like Epic's MyChart, it is pretty easy to export your blood work history, then add any medications you take and start building your own view. Sam admitted he gets a yearly sign-off from his doctor and then forgets about it, without tracking how his blood work trends year over year. He agreed that keeping it local makes sense.
Sam had been using the new Codex desktop app, which he says feels very similar to Cowork, with the GPT-5.2 Codex model on extra high reasoning. These were the versions available when we recorded in early February 2026. His take: Claude Code is more instantaneous, which makes it fun. Codex on extra high takes five to eight, sometimes ten minutes per run.
The wait paid off on polish. Sam took a product he had already built and brought it into Codex, and the deep reasoning found a ton of bugs and missing gaps he had not been paying attention to. We compare the two tools in more depth on our Claude Code vs OpenAI Codex page.
The product Sam polished is a day-by-day personal trainer that gives feedback after every set. In settings you choose a workout duration, a goal (his is to gain muscle mass), workout days (minimum two, he picked five), a body focus of upper, lower or balanced, an experience level, the equipment you have and any limitations.
During a workout you log each set. When Sam logged a set as a struggle but hit the target reps, it told him to keep going at the same rate. When he logged fewer reps and still marked it a struggle, it ran a fatigue check and offered to step down or end the exercise. He chose to step down, and it dropped the weight from 20 to 7.5 for the next set. The idea came from a ChatGPT thread he had been updating at the gym, which got annoying to type into. See the rep-by-rep AI personal trainer build page.
Sam's rule from the build: use a lighter reasoning model when you want quick iterations, and switch to deeper reasoning when you are polishing something and making it more production worthy. He found Codex a longer, less engaging way to develop, but worth it for that last stage. Jordan saw the trainer as part of a bigger pattern: it is now much easier to build software for yourself instead of reaching for an app or a spreadsheet.
The news that week, as of the February 5, 2026 recording, was Anthropic releasing Opus 4.6 with new agent features in Claude Code, OpenAI launching Frontier for enterprises to manage AI coworkers, and Perplexity launching a model council mode where a chair LLM synthesizes answers from several models. Jordan sees the labs reaching feature parity one launch at a time.
His bigger point was about how people use these tools. For most of the history of chatbots you asked one model something and got a response. Claude Code made work more sequential, with task lists. Now agents manage other agents.
I think the biggest change we're seeing now is multi agent management with other agents managing those agents. And I think that's gonna be the biggest real pivot here in using these LLMs is you become a manager of managers, and that manager is managing a bunch of other agents.
— Jordan Metzner, Episode 30
Sam said his own workflow over the previous few weeks had become managing a few agents at once, and he only has the capacity for three or four. He had not tried an agent management platform yet, but wondered what happens if one lets him manage 20.
But if it gives me the ability to, like, think through or effectively manage 20 different agents, I already feel, four x more productive with managing four.
— Sam Nadler, Episode 30
Software stocks had been hit that week. Jordan thinks it is still too early to tell whether companies like Salesforce and Workday adopt AI and lead, or whether new AI startups replace them, and expects winners and losers like every other cycle. Sam thinks there is some panic that will settle, but expects a few former giants to have a hard time within 24 months. Internally, he says, we are already building software for ourselves that we used to buy off the shelf, and the build or buy trade-off has changed.
Jordan walks through the DNA dashboard and Sam demos the trainer on screen. Watch Episode 30 of Built This Week, where the full transcript is also on the page. For more on AI and human biology, see Episode 32.
