There is a quiet, awkward moment happening inside a lot of companies right now. Someone builds an AI agent to handle a chore - triage the inbox, update the ticket, pull the report - and then, to make it work, they hand it a login. Not a scoped, careful, thought-through login. Usually a real one. The agent now has the keys, it never sleeps, and no one can quite say what it touched yesterday. Agentic Fabriq exists because of that moment.
The company, part of Y Combinator's Winter 2026 batch, describes itself in four words that a security engineer will immediately understand: Okta for agents. Okta answers a boring, load-bearing question for human employees - who are you, and what are you allowed to touch? Agentic Fabriq asks the same question about software that acts like a person but multiplies like a spreadsheet.
01The one rule
Strip away the diagrams and there is a single principle underneath the product: an agent should only ever act with the permissions of the user who invoked it. If Dana in finance can see three folders, the agent Dana asked for help should see three folders - not the whole drive, not the CEO's calendar, not the production database. Least privilege, an old idea in security, applied to a new and slippery kind of actor.
"Control what your agents can access and do - for every user."Agentic Fabriq
That sounds obvious until you try to build it. Human identity tools assume a person clicks a login and sits inside one session. Agents don't behave that way. One agent might act for a hundred different users, call a dozen tools, chain a request through three services, and do it all in a second. The permission has to travel with the request, get re-checked at each hop, and leave a mark. Agentic Fabriq's job is to sit in that middle and enforce it.
02What it actually checks
The clearest way to picture the platform is as a grid. On one axis, the agent and the user behind it. On the other, the resource being requested. Every call lands in a cell, and every cell has an answer. The policy engine evaluates that in real time, then either exchanges a token and lets the call through or stops it and writes down why.
| Agent × User | Inbox | Payroll | Prod DB | Calendar |
|---|---|---|---|---|
| Support bot / Dana | ALLOW | DENY | DENY | ALLOW |
| Finance bot / Dana | DENY | ALLOW | DENY | DENY |
| Support bot / Guest | DENY | DENY | DENY | DENY |
Around that engine sit the parts that make it usable. A credential vault so employees can connect their own accounts to tools without ever sharing a password or leaning on a broad service account. Prebuilt connectors for the usual suspects - Gmail, Google Drive and Calendar, Slack, GitHub, Notion, Microsoft 365 - so teams skip rebuilding authentication for every integration. And an audit trail that is signed, searchable, and immutable, plus one-click revocation for the day an agent starts behaving in a way no one intended.
Getting started is meant to be a few lines of code, not a quarter-long project. A team adds the SDK - TypeScript or Python - registers its agents and the tools those agents are allowed to reach, points the platform at an existing identity provider, and sets per-role policies. From there it deploys and watches. The five steps the company lays out read less like a security overhaul and more like wiring up a new dependency, which is the point: if governance is painful, people route around it, and then it protects no one.
The setup, in five steps03The founders
Agentic Fabriq was started in 2025 by Paulina Xu and Matthew Xu. They are not siblings who grew up scheming in a garage - they met at MIT admit weekend and became friends, then made the call to leave school before their second year to build this full time. Paulina, the CEO, came up through AI and physics, with machine-learning research at the MIT Kavli Institute, the MIT Haystack Observatory, and INAF Padua. Matthew left at 19 with a background spanning AI, math, and research stints at MIT's CSAIL and Futuretech Lab.
A pair of physics-and-math researchers decided the most interesting problem in AI was not the model. It was the plumbing around it.The bet behind Agentic Fabriq
That instinct shows in the product choices. This is not a flashy consumer app or another wrapper around a chatbot. It is infrastructure - the kind of thing that is invisible when it works and catastrophic when it doesn't. Choosing to build it takes a certain temperament: a comfort with being the unglamorous layer everyone depends on and no one talks about at dinner.
There is also a business logic to it. Infrastructure companies are hard to start and hard to kill. The first customer is the slow part; once a security team routes its agents through a control plane, ripping it out is its own project, and every new agent added afterward only deepens the dependency. Selling to the people who own identity and compliance means selling to buyers who think in years, not sprints. For a two-person team, that is a feature, not a limitation - it trades a fast, crowded consumer race for a slower one where being early and being trusted compounds.
04Why it is landing
One data point stands out. Agentic Fabriq reached a paying customer in roughly two months. Enterprise infrastructure usually takes twelve to eighteen, because buyers are cautious and the sales cycle is long. When money moves that fast, it is rarely about the pitch. It is because the pain is already sitting on someone's desk - a security lead who has watched agents multiply and realized there is no answer to the question "what can they touch, and how would we know?"
The timing helps. Every serious agent framework - LangChain, CrewAI, AutoGen, OpenAI Agents - has made it trivial to spin up something autonomous. The protocols underneath, MCP and OAuth, are still settling. That combination, cheap agents plus unsettled rules for how they authorize, is exactly the seam a company like this slots into.
05Who it competes with
The honest answer is that the biggest competitor is a shrug - the shared service account, the over-scoped API key, the homegrown token exchange that a team wrote in an afternoon and now cannot fully explain. On the vendor side, the human-identity giants like Okta and Microsoft Entra are the obvious incumbents, and they will not ignore agents forever. A wave of agent-native startups is forming around the same problem. Agentic Fabriq's wager is that identity for agents is different enough, and urgent enough, that a focused team gets there first.
Most security tools slow you down. The clever framing here is that governance can speed you up - because you stop rebuilding auth for every new integration.On the business logic
06Where it fits
Zoom out and the market is drawing a new layer of the stack. There are the models. There are the frameworks that turn models into agents. There are the tools agents reach for. And now there is the question of trust between all of them - who is allowed to do what, and where is the receipt. Agentic Fabriq is planting a flag on that last one. It is a small team making a large claim: that the agentic era needs its own Okta, and that it intends to be it.
Whether that claim holds is the open question of the next few years. But the shape of the problem is real, the early buyers are real, and the founders picked the boring, load-bearing part of the puzzle on purpose. In infrastructure, that is usually the tell worth watching.