There's a specific moment that stuck with us. We were watching a demo of an agent doing real work inside a SaaS product — clicking through dashboards, filling forms, looking genuinely useful. And then the question hit: would you trust that agent to do this at 3 AM, unsupervised, when it matters?
That's the question behind HIVE.
Agents are brilliant in demos and brittle in production. We kept seeing the same pattern: a screen-scraping agent works beautifully in a controlled demo, then breaks the moment the UI changes. A redesigned button, a moved field — and it fails silently, doing the wrong thing with full confidence. Every pixel the agent has to "see" has a price tag, in tokens, latency, and fragility.
But the deeper problem isn't fragility. It's blindness. An agent clicking through a UI can't tell safe from dangerous. It has no concept of permissions — it just does what the screen lets it do. Working in cybersecurity, we've seen what real access control looks like: permissions defined up front, inherited from the user, and an explicit yes before anything irreversible.
That became the thesis. Agents shouldn't scrape screens. They should operate inside a contract: structured access to a product's real capabilities, permission-controlled, with a human in the loop before high-impact actions.
We started in discovery — talking to founder-developers and engineers who are actually building AI products, and iterating on what we learned. We're not building this for enterprise buyers. We're building it for developers and startups shipping AI products, who need their agents to work in production the same way they work in the demo.
And the market is starting to agree. In late September, Genea shipped a native MCP server: their access-control software now lets agents manage doors and users in plain language — permissions off by default, inheriting the user's existing permissions, waiting for an explicit yes before high-impact actions like revoking access. The exact shape we argue for, controls built first.
One honest caveat: Genea owns their software, so they could build this first-party. HIVE's bet is the thousands of SaaS products whose vendors will never build this themselves. That's the gap we're closing.
We're building HIVE because the demo isn't the product. Production is.
What breaks first when your agents hit production?
Top comments (0)