Where Wrappers Go to Die

This is a scene from a demo day in Pangyo last spring. On stage, a founder demonstrated their company's agent — a workflow that reads meeting notes, schedules appointments, and drafts follow-up emails. The audience applauded. But during Q&A, one judge asked: "The model underneath — that's OpenAI, right? If they ship the same features as a default in six months, what's left of your company?" The founder paused for a moment. That silence was the real question facing Korea's agent market in 2026.

Over the past eighteen months, a wave of Korean companies have attached the label "agent" to their products — sales agents, legal agents, CS agents. Investment followed. But every gold rush has an end. When most miners return empty-handed, what survives are the ones who sold the pickaxes and jeans. In the current phase of separating wheat from chaff, the question to ask is not the cliché that "wrappers will die." It is something more specific: what becomes a moat above the model layer, and what evaporates?

What Evaporates and What Remains

First, what evaporates. Features that require a single model call are almost entirely gone: summarization, translation, tone adjustment, simple classification. These are not features — they are model defaults. "Wrappers" built through careful prompt engineering are equally at risk. Prompt engineering has near-zero replication cost. A competitor can screenshot your interface and reproduce the same output in a day. The moment a model company adds that functionality as an SDK default, your six months of work gets erased in a single line of release notes.

What remains is different. Four layers above the model hold onto margin. First: data. Anyone can call the model, but no one can summon accumulated data from a specific industry. Defect repair histories from Korean construction sites. Patient record patterns from veterinary clinics. Parts disassembly data from heavy-equipment yards. These are coordinates the model was never trained on. Second: workflow. Not a one-shot call, but a multi-step process with human checkpoints and rollback on failure — the business flow itself, embedded deep into a customer's operations. A customer who has wired this into their daily work does not walk away easily. Third: trust. In domains where errors trigger lawsuits and accountability must be traceable, "who stands behind this?" is a product in its own right. Fourth: regulation. Implementations that have cleared Korea's Medical Act, Personal Information Protection Act, and financial compliance requirements constitute a barrier to entry in themselves.

The Real Enemy Is Not the Founders

This is where a familiar diagnosis surfaces: "Korean founders don't bet as big as Americans," "the technology is there but the boldness isn't." I find this diagnosis lazy. The founder at that demo day was already looking at global markets. The problem was that he had no ecosystem beside him to help design which of the four — data, workflow, trust, or regulation — he should be building on top of the American model.

In the US, this design work is not outsourced. Firms like a16z don't just provide capital — they pattern-match "how AI apps survive model commoditization" and inject those patterns into their portfolios. YC conducts post-mortems after every batch on which layers evaporated, and passes those findings to the next cohort. The accumulation of these post-mortems is the ecosystem's intelligence. Korea has the capital, but it lacks this accumulated record. Founders keep falling into the same traps from scratch. Where a conversation should begin at its fiftieth iteration, it resets to the first every time.

The Bottleneck Is Customer Access, Not Capital

Capital, talent, regulation, customer access — which of these is the real bottleneck for Korea's agent startups? Capital is, surprisingly, less urgent. Seed-stage money is circulating. Talent is growing too; engineers who can work with LLMs are multiplying quickly. The true bottleneck lies in two places.

The first is customer access. Data and workflow moats only form when you are "deeply embedded in actual operations." But large Korean conglomerates and public institutions rarely open their operational data to early-stage startups. They ask for references, yet you need a first customer to build those references — a circular trap. While American startups infiltrate thousands of SMBs through self-serve channels and accumulate data in the process, Korean founders spend six months waiting for a single conglomerate to approve a proof-of-concept. In those six months, the model company ships that feature as a default.

The second is the pathway for converting regulatory moats into assets. Korea's rigorous regulations are not a burden — they can be a weapon. Implementations that have cleared medical, financial, and personal data requirements are credentials that carry weight globally. But every startup clears these regulatory hurdles alone. The ecosystem has not turned this hard-won experience into a shared asset. Regulatory sandboxes exist, but the lessons learned inside them do not flow to the next company.

What Korea Should Build on Top of American Models

The conditions under which Korean agent startups can protect their margins are clear. Use American models without apology — treat them as infrastructure. But build three things on top that America cannot access: closed-loop data from Korean industries, trust implementations that have cleared Korean regulations, and workflow standards for the Asian markets closest to Korea. This is why a B2B startup in Busan handling inbound inventory data from a Pyeongtaek heavy-equipment yard holds coordinates that Silicon Valley will never possess.

What the ecosystem must do is equally clear. Not spray more capital — accumulate post-mortems, build bridges to first customers, and turn regulatory clearance experience into a shared asset. Record which layers evaporated each quarter, and hand that record to the next founder.

There is a counterargument: "Even so, if model companies eventually integrate vertically and absorb industry-specific data, Korea's moats will collapse." It is a fair point. But model companies cannot go deep into every industry's ground-level operations simultaneously. Depth is always narrow and local. The day OpenAI directly collects Korean construction defect data will never arrive. That narrow, deep coordinate is the moat.

Founders already have an intuition for what to build on top of the model. The silence at that demo day was not ignorance — it was the absence of an ecosystem to design alongside them. Founders are already looking at the world. Now the ecosystem needs to catch up.