The founder building the firewall for AI agents
Wojciech Blaszak started selling software at 14 and walked away from university after a single semester. With Golf (YC X25), he is building the layer that lets a company finally see what its AI agents are doing with its data.
The uncomfortable fact at the center of Wojciech Blaszak's company is one that most enterprises would rather not say out loud. Somewhere inside their walls, an AI coding assistant is reading a private repository. A chat agent is pulling customer records to answer a question. A tool nobody officially approved is quietly connected to a database. And almost nobody can tell you exactly what is talking to what. That gap - the invisible space between an AI agent and the data it touches - is the whole reason Golf exists.
Blaszak is the co-founder and CEO of Golf, a San Francisco company that went through Y Combinator's Spring 2025 batch. The pitch fits on a bumper sticker: your agents connect to your data and you can't see it. The product underneath is more ambitious. Golf sits at the layer of the Model Context Protocol, or MCP, the emerging standard for how AI agents plug into tools and information. Instead of watching the language model, Golf watches the connections. It discovers every agent and data link in an organization, including the ones running in the shadows, enforces policies on them, and keeps an audit trail that a compliance officer can actually read.
If that sounds like a security company, that is because it has become one. But Golf did not start there, and the path it took says a lot about how Blaszak works.
01 / The teenager with a P&LBuilding before it was cool
Most founders can point to an early spark. Blaszak's is unusually early and unusually concrete. He started his first company at 14. Not a lemonade stand, not a school project - an actual business that he scaled to around $200,000 in revenue. For a teenager still years away from a driver's license, that is a full education in the parts of a startup that no class teaches: finding customers, charging them money, and delivering something they will pay for again.
The single semester of university is the other number that keeps coming up. Blaszak enrolled, stayed for one term, and left to build full-time. He is careful to frame it the right way. Dropping out was not the point. Having something better to do was. By the time he walked away from the lecture halls, the work that would become Golf was already alive and pulling on him harder than any degree could.
Before Golf took its current shape, he had already been in the trenches of applied AI. He worked on an AI SDR - the kind of automated sales agent that emails prospects and books meetings - and on infrastructure for AI agents more broadly. That is where the thread starts. Building agents teaches you very quickly that the hard part is rarely the intelligence. It is the plumbing.
02 / The co-founderTwo builders who started young
Blaszak did not build alone. Golf's co-founder and CTO is Antoni Gmitruk, a longtime collaborator with a matching origin story. Gmitruk built a fully automated 3D printer at 15 and, like Blaszak, left university after a single semester. The two are Polish, they had been building things together long before Golf, and they even co-founded a community for hackers and scientists back home.
There is a certain symmetry to it. One founder sold software as a teenager; the other automated hardware. Both learned the same lesson from opposite ends - that shipping something real, early, beats waiting for permission. When they pointed that shared instinct at the AI agent boom, they landed on the one part everyone else was treating as a chore.
Partnerships that start in adolescence tend to have a shorthand that new co-founders spend years trying to build. By the time Golf existed, Blaszak and Gmitruk had already tested each other on smaller projects, argued about how things should be built, and figured out who is better at what. That history matters more than it looks. Early-stage companies live and die on how fast the founding team can make decisions and recover from bad ones, and two people who have been iterating together since their teens do not have to relearn how to trust a call under pressure.
03 / From Authed to GolfThe pivot hiding in plain sight
Golf was not always called Golf. The company started life under the name Authed, and the name change tracks a real shift in what the product is for. The original instinct was about trust and authentication for AI agents - the identity layer. That instinct never went away. It grew up. As Blaszak and Gmitruk watched teams try to ship MCP servers, they saw the same story on repeat: two to four weeks of work on authentication, tracing, transports, and debugging before anything ever reached production.
So the first public version of Golf attacked that pain directly. GolfMCP is an open-source framework that lets a developer define the tools, prompts, and resources they want agents to call, and hands off everything else - routing, auth, telemetry, error reporting, deployment. The hosted Golf Gateway takes that further, deploying a server in under a minute with real-time traces and monitoring built in. What used to be a month of plumbing became a single deploy.
Open source got developers in the door. But the bigger question showed up the moment those servers went live inside real companies. Once agents are connected to your data, who is watching them? That question turned a developer tool into an enterprise problem, and the enterprise problem turned Golf into a security company.
04 / Below the model, not above itWhere the risk actually lives
Here is the strategic bet that makes Golf interesting. Most AI security tools point their attention at the model - the prompt box, the output, the language layer. Golf points its attention one layer down, at the connections. Blaszak's argument is that the real exposure is not in what the model says. It is in what the agent can reach: the repositories, the record systems, the credentials sitting behind a tool call.
The product breaks into three verbs. Discover: find every AI agent, MCP server, and data connection in the organization, including the shadow ones that traditional tools never see. Enforce: apply granular policies per tool, per team, and per data source, blocking prompt injections, PII leaks, and credential exposure with what the company describes as sub-millisecond latency. Audit: keep activity trails and pre-mapped evidence for the frameworks that keep security teams awake - SOC 2, ISO 27001, NIST AI RMF, FINRA.
It is a familiar shape if you have watched enterprise software before. Shadow IT was the last decade's headache - the unsanctioned apps and services employees adopted faster than IT could track. Shadow AI is the same story with higher stakes, because an ungoverned agent does not just leak a spreadsheet. It can act. Blaszak is betting that every large company will eventually need to see, control, and audit the agents it never officially approved, and he wants Golf to be the place they do it.
05 / The timelineOne year, a lot of ground
Compress that into roughly a year and the pace tells its own story. The company shipped a free framework, earned developer trust, spotted the enterprise problem underneath, and hardened into a governance platform with real compliance certifications - all while the standard it depends on was still being written in public. Building on top of a protocol that is itself in motion is a gamble. It is also exactly the kind of early, unglamorous bet that both founders have been making since they were teenagers.
06 / What he is actually buildingThe seatbelt for a car already moving
Talk to anyone shipping agents inside a big company right now and you hear the same tension. The tools are useful enough that teams adopt them immediately, and dangerous enough that security teams want to slam the brakes. Blaszak's answer is not to slow the car down. It is to make it safe to keep driving. Golf's whole design assumes agents are already loose in the organization and asks the practical questions instead: what are they touching, what should they be allowed to touch, and can you prove it later.
That framing shows up in the customers Golf points to. The company connects to more than forty integrations and the agent tools people actually use - Cursor, Claude Code, GitHub Copilot, ChatGPT Enterprise, Windsurf, and custom-built agents. One enterprise customer, quoted anonymously, put the value in blunt terms: Golf gave them governance for AI tools they do not control, and that was the actual problem nobody else was solving.
It helps that the framing lines up with how large organizations already think about risk. Security teams do not get budget for hypotheticals; they get budget for the things they can be audited on. By mapping agent activity to the exact compliance frameworks a company already lives under, Golf turns a fuzzy fear into a line item. That is a subtle but important move. It changes the conversation from "should we trust AI agents" to "here is the evidence that we control them," which is the version of the question a security leader can actually take to a board.
There is a version of the AI story that is all spectacle - bigger models, flashier demos, louder claims. Blaszak's version is quieter and, arguably, more durable. He has spent his short career finding the boring layer underneath the excitement and turning it into a business. A P&L at 14. Agent infrastructure before agents were a headline. And now the governance layer for a protocol that could sit under the next decade of enterprise software. The pattern is consistent enough that it stops looking like luck.
Golf is still young, and the market it is chasing is being invented in real time. But the bet is clear-eyed. If AI agents are going to run inside every serious company, someone has to build the layer that keeps them honest. Wojciech Blaszak decided that someone might as well be him - and, true to form, he started before most people noticed there was anything to build.