Week of Sep 11, 2026 Batch SaaS

Agent identity / human-proof gateway

Startup Idea Brief Published Week of Sep 11, 2026

A gateway that verifies whether an API caller is a human or an AI agent, so platforms can authenticate, price, and limit each differently.

AI agents now place orders, call APIs, and negotiate on behalf of people, but most authentication systems were built assuming every caller has a human behind the keyboard clicking a CAPTCHA. This company builds a gateway that sits in front of an API or checkout flow and answers a narrower question: is this request coming from a verified human, a verified agent acting on a specific person's or company's behalf, or an anonymous bot with no accountable owner behind it. Platforms plug it in the way they plug in a fraud-scoring API today.

The wedge is agent-aware identity and payments, not generic bot blocking. Existing bot-mitigation tools are built to stop traffic; this product is built to let legitimate agent traffic through with a verifiable chain back to a paying, liable human or company, while still stopping traffic with no such chain. That distinction matters because platforms increasingly want agents to transact, not just be filtered out, and need a way to attach fraud liability, spending limits, and an audit trail to a non-human actor.

Customers are marketplaces, fintechs, and API-first platforms that are starting to see agent traffic in their logs and don't yet have a policy for it, let alone infrastructure to enforce one.

Why Now

Agentic AI moved from demos to production in 2026: agents now call checkout APIs, book services, and negotiate B2B deals with no human clicking through each step, which breaks authentication and fraud models built around a person at a keyboard. That shift creates real liability exposure, since a platform that can't tell a verified agent from a scripted bot has no way to price risk, enforce spending limits, or hold anyone accountable when something goes wrong. Regulators and standards bodies are only beginning to define what an agent's legal identity even is, which means the platforms and vendors active in that conversation now will help shape the rules rather than inherit them later.

Demand Signal

Y Combinator's own Fall 2026 Requests for Startups includes a category built around this exact problem, Proving You're Human, framed by YC partner Max Kolysh around cases like the 2024 Hong Kong deepfake fraud in which an employee wired 25.6 million dollars after a video call where every other participant turned out to be fake. Two YC-backed companies are already building pieces of this stack: OneCLI positions itself as an identity gateway that binds every AI agent to a specific employee's identity with a logged audit trail, and Clawvisor built an authorization layer that vaults credentials and requires human approval before an agent can act on connected services like payments, email, and calendars. A third YC company, Roundtable, is attacking the adjacent problem of verifying humans against bots with an API called Proof of Human. That much funded activity and a dedicated YC category in the same twelve months signals this is being treated as a near-term infrastructure gap.

Defensibility

The moat compounds through a trust network effect: once enough platforms rely on the same identity gateway to vouch for agent traffic, agents and their operators need to register with it to get through anywhere, which pulls in more platforms in turn, similar to how early fraud-scoring networks became infrastructure. Active participation in the technical standards that don't exist yet, for how an agent proves it acts on behalf of a real accountable party, lets an early mover help write the interfaces everyone else has to integrate against. And as agent-identity regulation gets drafted, a vendor with production data and a seat in that process is positioned to become a reference implementation regulators point to, a position a later entrant would find much harder to take.

Market Signals

YC Fall 2026 RFS category: Y Combinator's Fall 2026 Requests for Startups lists Proving You're Human as its own category, citing the 2024 Hong Kong deepfake video-call fraud that cost a company 25.6 million dollars. Y Combinator

OneCLI funded by YC: OneCLI, a YC Summer 2026 company, markets itself as an identity gateway for AI agents that binds every agent to an employee's identity with a logged audit trail. Y Combinator

Clawvisor funded by YC: Clawvisor, a YC-backed company, built an authorization gateway that vaults credentials and requires human approval before an agent can act on connected payment, email, and calendar accounts. Y Combinator

Adjacent funded competitor: Roundtable, another YC-backed company founded by two Princeton cognitive scientists, built Proof of Human, an API that detects bots and verifies human identity, showing venture interest in the verification side of the same broader problem. Y Combinator

How I'd Build This

MVP Scope

Ship a single API endpoint and SDK that platforms drop into their checkout or agent-facing API layer: it takes a request, checks it against a registry of verified agent credentials plus behavioral signals, and returns a verified-human, verified-agent, or unverified label with a confidence score and the calling party's identity when known. No dashboard or policy engine yet, just the classification call and a webhook for logging.

First Customers

Target the ten or so API-first fintechs and marketplaces already seeing agent traffic today, found by watching who is publicly discussing agentic checkout, agent payment protocols, or bot fraud on their own engineering blogs and at agent-commerce events, and by approaching companies rolling out agent payment standards. Offer free integration and a shared case study in exchange for being a design partner while the classification model is still being tuned.

Moat Sequence

Start with the classification API to get real traffic and label agents accurately, then add the credential and audit-trail registry once several platforms are sending it identity claims, and only after that build the cross-platform trust network where one platform's verified-agent history carries weight with the next platform. Each layer needs the traffic volume from the one before it to be worth building.

90-Day Timeline

First 90 days: build and pilot the classification endpoint with three design-partner platforms and measure the false-positive rate on real agent traffic. Months 3 to 6: add the credential registry and pursue at least one standards or working-group engagement.

Sources

OneCLI | Y Combinator Clawvisor | Y Combinator Requests for Startups | Y Combinator Roundtable | Y Combinator