From Vibe Coding to Agentic Engineering: Everyone Can Now Build for Healthcare
The most consequential shift in software is not about better developers. It is about removing the need to be one.
Key Takeaways
- Andrej Karpathy coined vibe coding in February 2025; by February 2026 he declared it passé, replaced by agentic engineering—AI-assisted development with professional oversight and scrutiny
- At Anthropic, 70–90% of production code is now AI-generated; Claude Code accounts for ~4% of all public GitHub commits globally
- OpenAI acquired Windsurf ($3B) and Torch ($100M health AI)—a deliberate bet that domain experts with AI tools will build the next wave of healthcare software
- India’s NHCX mandates 1-hour pre-authorization and 3-hour discharge approvals; most claims currently take 15–30 days—that gap is the opportunity
- The “one rail, many claims” architecture of NHCX means any solution built on its APIs can reach 500+ million ABHA-registered citizens from day one
1. A Throwaway Tweet That Changed Everything
On a quiet February evening in 2025, Andrej Karpathy—co-founder of OpenAI, former head of AI at Tesla—fired off a stream of thoughts on X. He called it “vibe coding”:
“There’s a new kind of coding I call ‘vibe coding’, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It’s possible because the LLMs are getting too good.”
He described barely touching the keyboard. He used voice input to say things like “decrease the padding on the sidebar by half.” He clicked “Accept All” without reading the diffs. When errors appeared, he pasted them back in with no comment and waited for the model to self-correct.
The post was viewed over 4.5 million times. Merriam-Webster listed it as slang in March 2025. Collins English Dictionary named it Word of the Year.
But Karpathy himself moved on quickly. By February 2026—exactly one year later—he declared vibe coding was already passé. The new term: Agentic Engineering.
“Programming via LLM agents is increasingly becoming the default workflow for professionals, except with more oversight and scrutiny… ‘Engineering’ to emphasize that there is an art, science, and expertise to it.”
The distinction matters enormously for healthcare. Vibe coding was thrilling and chaotic. Agentic engineering is disciplined, structured, and—for the first time—safe enough to discuss seriously in clinical contexts.
2. The Numbers Behind the Shift
When Boris Cherny, the creator of Claude Code, shared his personal workflow in December 2025, the industry stopped and listened. The stats were jaw-dropping:
- 259 pull requests merged in a single month
- 497 commits, 40,000 lines added, 38,000 lines removed
- Every single line written by Claude Code + Opus 4.5
- Claude ran autonomously for minutes, hours, and sometimes days at a time
He noted: “Software engineering is changing, and we are entering a new period in coding history.”
This is not a hobbyist anecdote. At Anthropic, between 70% and 90% of all production code is now AI-generated. Claude Code itself represents approximately 4% of all public GitHub commits globally. Daily active users doubled in a single month.
The platform did not start as a grand vision. Cherny built the first version as a side project in September 2024. Less than 18 months later, it is the tool of choice for engineers at some of the world’s most demanding organizations.
What strikes us at Yajur Healthcare is not just the productivity gain. It is what these platforms do for people who are not engineers.
3. Big Tech Has Already Connected the Dots
Analyst Aakash Gupta drew a sharp observation about OpenAI’s acquisition pattern in early 2026:
“Altman just told you OpenAI’s acquisition strategy and nobody is connecting the dots. 10 acquisitions in the last year.”
Two acquisitions stood out for their strategic pairing:
- $3 billion for Windsurf — an AI-native coding IDE
- $100 million for Torch — a healthcare AI company
These are not coincidental investments. They represent a calculated bet that the next wave of healthcare software will not be written by hospital IT teams working off decade-old specifications—it will be generated by domain experts who understand the problem, using tools that handle the implementation.
The implications for healthcare are significant. When the friction of software development collapses, the people closest to clinical problems become the most capable problem-solvers. A doctor who has spent twenty years watching patients fail to follow up on referrals does not need a six-month development sprint to prototype a solution. She can describe it, and an agentic system can build it.
4. What This Looks Like in Practice
The Replit CEO shared a case that has circulated widely in healthcare technology circles: a British doctor built a working health-tracking application for his patients for approximately £200 (~₹21,000). No technical co-founder. No agency. No months of back-and-forth requirements documentation.
This is the pattern that matters. Non-technical domain experts—physicians, nurses, pharmacists, hospital administrators, insurance case managers—are beginning to build tools directly calibrated to their workflows. The outcomes of this shift are predictable:
Patient engagement tools built by the very clinicians who understand why patients disengage.
Referral tracking systems designed by coordinators who manually chase referrals today and know exactly where the process breaks.
Medication adherence apps created by pharmacists who have watched specific patient populations struggle with specific drug regimens.
Triage decision aids prototyped by emergency medicine physicians who have already mapped the decision tree in their heads.
Claims pre-authorization workflows designed by revenue cycle staff who know every payer’s quirk.
None of these require a computer science degree. They require clinical and operational expertise—which already exists in abundance across the healthcare ecosystem. What was missing was the tool to translate that expertise into functional software. That tool now exists.
5. India’s Own Answer: The NHCX Hackathon
On February 18, 2025, the National Health Authority launched its NHCX Hackathon Masterclass—a direct signal that India’s government-backed health infrastructure is ready to be built upon by anyone with a problem worth solving, not just established software vendors.
The National Health Claims Exchange (NHCX) is India’s answer to the fragmented, paper-intensive health insurance claims process. The problem it addresses is vivid: a hospital billing team treating three patients under three different insurance schemes must today log into three separate portals, fill three different claim formats, and track status across three different systems—for the same treatment, delivered by the same doctor. The NHCX core architect’s framing at the masterclass was precise: “one rail, many claims.” Think of Indian Railways—thousands of trains, different destinations, different speeds, but all running on the same infrastructure. Hospital talks to one NHCX. NHCX talks to every insurer. Built on FHIR standards and integrated with ABDM, every step of the claims lifecycle—coverage verification, pre-authorization, claim submission, adjudication, and settlement—travels on that single standardized rail.
The hackathon masterclass invited a new generation of builders to tackle the unsolved problems in this ecosystem:
Pre-authorization bottlenecks. Regulators mandate that cashless pre-authorizations be processed within 1 hour and discharge approvals within 3 hours—from the moment the first document is received. Yet today most claims cycle through 15 to 30 days of back-and-forth queries before settlement arrives. The gap between regulatory intent and operational reality is where the largest efficiency opportunity lives. The NHCX API exposes the pre-auth workflow programmatically, making it possible for the first time to build automated pre-auth engines without constructing the underlying data exchange layer from scratch.
Claims adjudication transparency. Hospitals submit claims into a black box and receive responses without structured reasoning. Builders on NHCX can create adjudication co-pilots that surface rejection reasons, suggest documentation remediation, and predict settlement timelines—tools that revenue cycle managers have wanted for years but could not commission at viable cost.
Fraud pattern detection. Standardized, machine-readable claims data flowing through NHCX creates, for the first time, a dataset rich enough to train fraud detection models at scale. A health data scientist or even a domain-expert clinician using an AI coding platform can prototype a claims anomaly detector in days.
Insurance policy digitization. Today, insurance policies arrive as long PDF documents that no billing team has time to fully interpret. A hospital guesses what is covered, submits the claim, and learns of exclusions only after adjudication—a source of enormous friction for patients and providers alike. NHCX exposes a structured InsurancePlan API that allows hospital systems to pre-check coverage programmatically before a procedure begins, moving the entire ecosystem from assumption-based to rule-based care delivery.
Real-time tracking for patients. Perhaps the most underappreciated opportunity: consumers have no visibility into where their claim stands. A claims status layer built on NHCX APIs—something a non-technical hospital administrator could prototype using today’s AI coding platforms—would eliminate the single most common complaint in health insurance interactions.
The masterclass represents something significant: a government infrastructure owner publicly inviting domain experts—not just IT companies—to build on an open platform. This is the ABDM and NHCX ecosystem making the same bet that the broader AI coding revolution has already validated: that the people closest to the problem are, with the right tools, the most capable builders of its solution.
And the numbers make urgency unavoidable. India processes hundreds of millions of health insurance transactions annually across 500+ million ABHA-registered citizens, 30,000+ hospitals, and over 200 insurers and TPAs. A pre-authorization delay that costs an average patient 6 hours is not a minor inefficiency at this scale—it is a systemic failure affecting millions of care episodes every year. A fraud pattern that costs 2% of claim value is not a rounding error—it is thousands of crores diverted from care delivery.
These problems cannot wait for traditional enterprise software procurement cycles of 18 to 36 months. They require solutions that can be conceived, built, validated, and deployed in weeks—which is now possible for the first time.
6. The Yajur Healthcare Perspective: Opportunity Meets Obligation
We have been thinking carefully about what this shift means for the infrastructure layer beneath healthcare software.
Enthusiasm is warranted. When a district-level nurse coordinator can prototype a home visit scheduling tool in an afternoon, and that tool integrates with ABDM to pull verified patient identifiers and push structured care records—that is transformative for primary healthcare delivery in India.
But the clinical AI failures we have documented in our earlier work are instructive. Models produce plausible code faster than teams can validate it. In one real-world deployment, an LLM misread a clinical eligibility rule and misclassified “no steroid use in the last six weeks” as “recent steroid use stopped five days ago”—an error with direct patient safety implications.
The lesson is not that AI-assisted development is unsafe. The lesson is that the infrastructure around the AI must be engineered with the same care as the model itself.
For healthcare specifically, this means:
Data validation must be schema-aware. A medication dose field that accepts free text is a liability. FHIR-compliant structured outputs are not a nice-to-have—they are the foundation on which clinical trust is built.
Provenance must be traceable. Every recommendation, every auto-populated field, every generated summary must carry a pointer back to the source. Auditability is not a compliance checkbox. It is how clinicians maintain oversight of AI-generated outputs.
Ambiguity must surface, not be hidden. Agentic systems that quietly guess when data is missing are dangerous in clinical contexts. The pipeline must flag uncertainty and escalate appropriately.
Human-in-the-loop is a design principle, not a fallback. The most effective deployments we have seen treat clinician review not as an exception path but as the default confirmation step for consequential decisions.
This is the gap that Yajur Healthcare exists to address. As coding platforms lower the barrier to building, we focus on raising the quality of the clinical context those tools operate within—structured pipelines, validated outputs, contextual AI that understands the difference between a lab value from this morning and one from three years ago.
7. India’s Compounding Advantage
The timing of this shift intersects with a unique moment for Indian healthcare.
The ABDM network is maturing. NHCX is beginning to digitize the claims ecosystem. UHI is creating a discovery layer for health services. These are not isolated pilots—they are network-scale infrastructure investments that, for the first time, give any developer—technical or clinical—a common data fabric to build on.
India also has an extraordinary pipeline of STEM graduates, multilingual clinical annotators, and domain experts across oncology, primary care, mental health, and chronic disease management who understand the local care delivery context in ways that no offshore team ever will.
When you combine:
- AI coding platforms that enable domain experts to build
- National health infrastructure (ABDM, NHCX, UHI) that provides structured interoperability
- A large, multilingual clinical workforce that understands local context
- A data infrastructure company (Yajur Healthcare) that engineers the clinical reasoning layer
You get something genuinely new: a healthcare software ecosystem that is built from the inside out—by the people who deliver care—rather than from the outside in, by technologists who must be taught what care actually looks like.
The speed at which this can now happen is the most underappreciated part of the story. Healthcare IT has historically operated on timescales that are wildly out of sync with clinical urgency. A hospital administrator who identifies a workflow breakdown today would, in the old model, spend six months writing a requirements document, another six months in vendor evaluation, and a year in implementation—by which time the context has changed, the champion has moved on, and the tool no longer fits. With AI-assisted development, the same administrator can prototype a working solution this week, validate it with real users next week, and iterate it into something deployable within a month. The feedback loop between problem and solution has compressed by an order of magnitude. At India’s scale—where every efficiency gain multiplies across a billion interactions—this compression is not incremental. It is transformational.
8. What We Are Watching
The platforms will continue to improve. Karpathy’s evolution from vibe coding to agentic engineering within a single year signals how fast the paradigm is moving. Boris Cherny’s prediction that software engineering as a title will transform—not disappear—points toward a future where clinical and technical expertise merge in practitioners we don’t yet have names for.
We are paying close attention to:
- Healthcare-specific coding environments that understand FHIR schemas, ICD-10 hierarchies, and ABDM APIs natively
- Agentic workflows that can handle the multi-step complexity of clinical processes—prior authorization, care coordination, discharge planning—not just isolated data entry
- Evaluation frameworks for AI-generated healthcare software that go beyond functional testing to include clinical safety, data privacy (DPDP Act compliance), and interoperability validation
- The regulatory posture of agencies like NHA and CDSCO as clinician-built software begins entering procurement pipelines
Frequently Asked Questions
What is vibe coding? Vibe coding is a term coined by Andrej Karpathy in February 2025 to describe AI-assisted software development where the developer describes intent in natural language and accepts AI-generated code with minimal review—sometimes without reading the diffs at all. It was named Collins English Dictionary’s Word of the Year for 2025.
What is agentic engineering? Agentic engineering is Karpathy’s follow-up concept, introduced in February 2026, describing a more disciplined form of AI-assisted development where LLM agents handle implementation autonomously while the developer maintains oversight, scrutiny, and engineering judgment. It is the evolution of vibe coding for professional and production contexts.
What is NHCX (National Health Claims Exchange)? NHCX is India’s government-built interoperability platform for health insurance claims, developed by the National Health Authority (NHA). Built on FHIR R4 standards and integrated with ABDM, it provides a single API framework—”one rail, many claims”—through which hospitals, insurers, TPAs, and government schemes exchange structured claims data. It eliminates the need for hospitals to manage separate integrations with each insurer.
Can non-technical healthcare professionals build software with AI coding tools? Yes. Platforms like Claude Code, Cursor, and Windsurf allow clinicians, administrators, pharmacists, and other domain experts to prototype functional software by describing what they need in plain language. A British doctor built a working patient health-tracking application for approximately £200 (~₹21,000) with no technical co-founder or agency.
What were the NHCX Hackathon 2025 problem statements? The NHCX Hackathon Masterclass (February 18, 2025) defined four problem areas: (1) converting legacy HMIS data into FHIR-compliant NHCX/ABDM format, (2) extracting data from PDF medical documents (lab reports, discharge summaries) into ABDM-compliant FHIR bundles, (3) converting PDF insurance policy documents into machine-readable FHIR InsurancePlan objects, and (4) fraud detection and claims processing time optimization using NHCX’s structured data.
What is the regulatory mandate for cashless claim processing in India? Under current IRDAI guidelines, cashless pre-authorizations must be processed within 1 hour and discharge approvals within 3 hours from the time the first document is received. Most claims today take 15 to 30 days due to repeated manual queries and fragmented workflows—a gap that NHCX is designed to close.
How does NHCX relate to ABDM? NHCX is built on the ABDM (Ayushman Bharat Digital Mission) ecosystem. Participants must be ABDM-registered to access NHCX APIs. ABHA IDs (Ayushman Bharat Health Accounts) are used to validate beneficiaries, and ABDM’s access control and security layer governs all NHCX transactions.
A Closing Thought
Healthcare has always suffered from a translation problem. The people who know what patients need cannot build the tools. The people who can build the tools cannot know what patients need.
AI coding platforms are dissolving that barrier. But dissolving the translation problem only creates a new obligation: the solutions that emerge must be worthy of the scale they are about to reach.
India does not have the luxury of slow. With 1.4 billion people, a rapidly aging population, a chronic disease burden accelerating faster than care infrastructure can absorb it, and a health insurance system that still leaves the majority of citizens navigating paper-based chaos—the gap between the problem and the solution has always been measured in lives, not just efficiency metrics.
What has changed is that the speed of building can now match the urgency of the problem. A clinician who sees what is broken can now build what fixes it—this week, not this decade. An infrastructure like NHCX that was built for scale can now be built upon at speed. A country with one of the largest skilled clinical and technical workforces in the world can now deploy that workforce directly on the problems it understands best.
The tools are ready. The infrastructure is ready. The builders—many of them not yet thinking of themselves as builders—are ready.
The question is whether we move with the urgency the moment demands.
Manish Sharma is the Founder & Director of Yajur Healthcare, a medical data infrastructure company building the context and annotation layer for safe clinical AI in India.
Connect with us if you are working on ABDM-integrated applications, clinical agentic workflows, or healthcare-specific AI tooling: connect@yajur.ai
This article was generated with the help of AI agents.