How to Design for Human-agent Interaction

Karri Saarinen, CEO of Linear (and former principal designer at Airbnb and Coinbase), argues that the unreliability people experience with AI products is fundamentally a design problem, not a model problem. When an agent sends an email you meant to review first, the failure was in the interface, not the model.

Saarinen presents a six-principle framework Linear developed for human-agent collaboration inside the same product. His core insight: non-deterministic software breaks the traditional design contract where a button does the same thing every time. When outcomes vary based on input, designing for reliability requires new patterns — and that responsibility belongs to designers, not researchers.

The piece also explores accountability when agents act incorrectly, a question Saarinen approaches with nuance. The article is based on his AI & I podcast appearance with Dan Shipper, available in video and audio formats.

RDCO mapping: Directly relevant to how we design agent interactions across our own stack. The principle that interface design determines trust more than model capability applies to our Notion/Slack/MCP integrations. The accountability framing is useful for the Sanity Check newsletter — the question of who is responsible when an AI agent acts wrong is underexplored territory. Sponsor note: Dialect (Scale AI) ad present — external sponsor.