Research at Your Cursor: Building a Proactive Insights Agent on Superhuman Go

Don’t search for research. Let the right insights find you.

Every company eventually encounters the same problem: it has more research insights than people can reasonably remember.

A product manager writing a strategy may make a claim about what customers need while three existing studies complicate that assumption. The evidence exists, but the author may not know it exists or where to find it. Looking for it means interrupting the work, opening a repository, searching through reports, and determining what still applies. Too often, the moment passes without the evidence that could have strengthened the strategy.

The traditional solution is to build a better research repository. That solves part of the problem, but it still depends on people recognizing that they need research, leaving their workflow, and knowing what to ask.

Inspired by our proactive AI assistant, Superhuman Go, we began exploring a different model: What if research could show up at your cursor?

So we built our Interactive Research Insights System, or IRIS. But turning a repository into a proactive assistant required us to build two distinct components: a connector and an agent.

A connector for access. An agent for action.

First, we built an MCP: a reusable service that lets compatible AI systems access our curated body of UX and marketing research.

Within Go, that MCP becomes a connector, an integration between Superhuman Go and a third-party tool (like Glean or Slack) that gives agents secure access to an external tool or knowledge source.

Our connector opens the door to the research repository, allowing Go to find relevant studies, retrieve evidence, and link back to original sources. Because the underlying service is portable, it can also support experiences in Slack and other compatible AI environments.

Then we used Go’s Agent Builder to create the IRIS agent. Unlike a connector, an agent has instructions, an interface, and triggers that determine how and when it should act. We connected the agent to our research connector, defined how it should use the evidence responsibly, and gave it two ways to help.

In chat, the IRIS agent works like a familiar research repository. Someone can open IRIS inside Go and ask, “What have we learned about how much control people expect when an AI agent takes action?” The agent uses the connector to search our research, synthesize the available evidence, and provide links to the relevant studies.

But we did not want people to benefit from IRIS only when they remembered to open it.

Using Go’s “while writing” trigger, we also enabled proactive underlines. The same agent can leave its chat window and meet people at their cursor.

The connector makes our research portable anywhere. The agent surfaces insights everywhere.

From a repository you visit to insights that find you

Most research repositories depend on pull: someone recognizes that they need evidence, visits the repository, formulates a question, and searches for an answer. Even when the repository works perfectly, the person still has to interrupt what they are doing.

Go unlocks a push experience. As someone at Superhuman writes, IRIS can recognize a potentially research-relevant claim and surface a proactive underline indicating that supporting, qualifying, or contradictory evidence may exist.

Inside Superhuman, cross-functional partners already use the IRIS chat experience to gather context and bring prior research into their work. Proactive underlines extend that value to moments when someone may not know what to ask, or may not realize relevant research exists.

The next time someone writes a product brief, strategy, or marketing plan, IRIS can make that connection in the moment. The writer can inspect the surfaced evidence via underlines, follow the original source, strengthen the claim, qualify it, or reconsider it … without leaving their work to begin a search.

A traditional repository provides a place to find research. By combining a connector, an agent, and Go’s proactive triggers, we created a way for research to find them.

An opinionated research agent

With Go, giving an agent access to research is easy; ensuring it represents that research responsibly is harder.

We designed IRIS to acknowledge when evidence is incomplete, limited to a particular audience, or insufficient to support a conclusion. It links people to original studies and recommends involving a researcher when the stakes are high, or uncertainty is high.

IRIS is deliberately more Socratic than authoritative. It can retrieve evidence, connect findings, and surface questions someone may not have considered, but it should not do their thinking for them.

IRIS democratizes access to existing research, not the expertise required to produce or interpret it. Researchers still frame the right problems, understand how evidence was produced, recognize consequential limitations, and judge how far a finding can reasonably take us. IRIS makes more of that evidence and recorded context accessible; people remain responsible for deciding what it means.

Building those boundaries into IRIS helped us see that becoming AI-native is about encoding the research team’s judgment into the systems we build.

What AI-native research means to us

Our team already uses AI to develop study plans and discussion guides, write analysis scripts, interrogate data, and communicate findings. Those practices make parts of research faster.

Building IRIS taught us to think beyond individual research tasks. We also build programs, systems, and infrastructure. Becoming AI-native means asking which parts of our collective knowledge should become reusable capabilities, where human judgment must remain explicit, and how AI can connect people rather than create another private workflow.

This work reflects our commitment to being (Super)human-centered, one of our company values. Employees should not need to be familiar with the research team’s repository to benefit from what users have taught us. At the same time, the technology should bring evidence and people into closer conversation and not replace that conversation.

We began by building a place where people could ask about research. With Go, we built something that can come to them: research insights available at the cursor where decisions take shape.

A note of gratitude

Special thanks to Colin Burke, Staff Quantitative Researcher, whose collaboration, technical imagination, and work building IRIS’s infrastructure made this project possible. It also exemplifies how researchers can help teams become truly AI-native. I’m equally grateful to our entire research team, who continually show up with curiosity, willingness to experiment, and the discernment to evolve our craft without losing what makes it valuable.