Your idea is now open source
The idea is worthless, execution is everything. My own rule of thumb is harsher. Whatever idea you have, assume 8,000 to 10,000 other people on Earth have it too. What you do with it is the only part that's yours.
Building has been getting cheap for a while: vibe coding turns a description into a running app in an afternoon. This month copying got cheap too. Below: what's left of execution when both are cheap, and what that means for healthcare, where the hardest work never lived in the code.
Building and copying, both in minutes
REA, short for Reverse Engineer Anything, is a free, open-source tool for coding agents: point it at an app or website, without source code, and ask in plain English how a feature works. It gained around 208,000 GitHub stars in a single week and now sits close to 400,000. Replit CEO Amjad Masad wrote: “Pretty soon all software will be de facto open-source.”
Writer Aakash Gupta pointed REA at Signal's desktop app and it answered all five of his questions. He then handed Codex, OpenAI's coding agent, a one-page spec of Linear, the $2.5 billion issue tracker, and had a working board three minutes and 28 seconds later, without Codex seeing Linear's code. On 15 October decomp.dev, a community that spent years decompiling video games by hand, shut down because AI had made the work too easy.
So if both are cheap, what does execution still mean?
The wrapper is the cheap part
Healthcare offered an answer on 22 October. Microsoft published a study in Science showing that MAI-DxO, a panel of AI agents that question a simulated patient, order tests and debate the diagnosis, solved 83% of complex cases, against 36% for generalist doctors (specialists weren't compared). And as TechCrunch pointed out, much of the gain comes from the models underneath: GPT-5 on its own scored 78.9%, and Microsoft's orchestration added about four points.
The orchestration layer, the conductor passing the case between agents, is the part an agent can rebuild, or you can buy off the shelf. The agents inside it are a different matter. A diagnostic agent that works in Boston has to be tuned again for a Belgian geriatric ward: to that department's protocols, the local rules on who may decide what, and the preferences of the clinicians and patients who use it. That takes data, feedback and months of work. What Microsoft can't ship from a repository at all is the next step. In the emergency department of Beth Israel Deaconess in Boston, consenting patients will also go through DxO, and its calls will be compared with the doctors'. A competitor would need years to copy that trial.
Even the dull plumbing is turning into code. At Charité in Berlin, researchers (one of them a Microsoft employee) wrote the whole hospital-governed cloud setup, from secure backend to access rights, into machine-readable files that are reviewed once and deployed identically anywhere. A compliant clinical AI service went live in 310 seconds on average. Their first tool, an inconsistency checker, flagged a severe inconsistency in about one in 13 physician-approved documents. Five minutes got that service running. Getting doctors to act on its warnings took longer: physicians checked 150 of its findings by hand, and 149 held up. That audit is what makes a ward willing to rely on the tool, and no agent can produce it from the outside.
How I'd split a healthcare product today:
Copyable in an afternoon
- The interface and the feature list
- The workflow logic behind a booking or triage screen
- The orchestration around a frontier model
- Increasingly, even the compliant cloud setup
Takes three years
- A trial with consenting patients and an outcome that holds up
- A CE mark under the EU's Medical Device Regulation (MDR)
- A reimbursement code, and a payer who believes your numbers
- Agents tuned to one department's protocols, rules and people
- A head nurse who vouches for you to the next ward
Germany shows what the second list costs. Of the 74 apps listed for prescription under its DiGA programme since 2020, only 14 could prove a benefit when they were admitted, per the statutory insurers' own report. Sixteen have since been delisted and nine makers went insolvent. For medical devices, notified bodies report 13 to 18 months for an average MDR review. None of this lives in the app, so none of it can be reverse engineered.
When building and copying both take minutes, execution becomes everything an agent can't reverse engineer: the evidence, the approval, the payer and the clinicians who vouch for you.
Where the hours should go
That's uncomfortable for founders, who spend most of their hours on the first list because it's easy to demo, while a competitor can now rebuild it over a weekend. For hospitals it's leverage they rarely use: trial sites and data access are the scarce inputs, and whoever offers them early shapes what gets built. And clinicians matter more than ever: their willingness to test, audit and recommend a tool is what no repository holds.
Try this on your roadmap this week. Mark every item an agent could rebuild from your public website and documentation by Friday. Then add up how much of last quarter's work went into the items you didn't mark. If the answer is "not much", that's where those 8,000 twins of your idea will overtake you.
💥 May this inspire you to spend your best hours on the part nobody can copy.