Mike Krieger on Building Agent-Native and Malleable Software

Key Takeaways

• Software has moved past the era of bolted-on AI sidebars and mini-windows that answer isolated support questions.
• True agent-native design requires shared plumbing where an AI agent can execute any action a human user can perform via REST APIs.
• At Anthropic, Claude dynamically builds, maintains, and updates internal tracking interfaces as organizational priorities shift.
• Unlocking emergent agent behavior depends on building infrastructural primitives rather than static, one-off UI features.

The Death of the Disconnected AI Sidebar

Most product teams slapped a chat box on the right rail of their web app and declared their product AI-enabled. Mike Krieger, product leader at Anthropic and co-founder of Instagram, points out why that approach hit a wall. As Krieger put it: “V1 or like er 1 was great, we've got this AI thing, maybe it belongs in like a sidebar or like a little mini window and you can ask questions and it's kind of disconnected and if you know maybe it helps you with some customer support questions.”▶ 12:57

That sidebar model fails because the agent cannot touch the product core. It sits outside the workflow, reading static text and offering generic suggestions. When an AI cannot click the buttons, update the records, or trigger the mutations that a human user can, it remains a novelty.

To break out of that trap, products need shared plumbing under the hood. Krieger tests new products by probing their infrastructural boundaries: “I'll ask it to do something that you know a human could do but you know might be required somebody to think through or architect the thing where everything goes through some shared plumbing that like both agents and like whatever sort of rest API you've built can go through as well.”▶ 16:01

Malleable Interfaces Built and Run by Claude

The next phase goes beyond parity between human and agent capabilities. The user interface itself becomes fluid. Instead of engineers hardcoding every dashboard and filter state, agents generate and reshape interfaces on the fly.

Krieger observed this shift firsthand inside Anthropic over the past year. “One of the biggest shifts in how we've worked in the last year I think at Anthropic is so much of the software we interact with is like built maintained and iterated on by claude,”▶ 14:42 Krieger explained. “Of course reacting to all the things that are happening in the organization but very much like changing all the time.”▶ 14:50

When an interface is malleable, software stops being a rigid set of pre-baked screens. The agent inspects live data, understands the project context, and creates the exact view a team needs for that week's sprint. Once the project shifts, Claude rewrites the UI components to reflect the new state.

Building for this world requires treating agent access as a first-class architectural requirement. Krieger noted: “it was the idea of like everything a human could do an agent should be able to do as well and very few products actually get this right. I think a lot of our products like don't quite get this right yet. But when you do that, all of a sudden sort of emergent behavior gets unlocked.”▶ 13:29

What to Do With This

Audit your core product's REST API this week. List the top ten actions human users take in your UI, from updating a pipeline status to exporting custom reports. If an external agent cannot execute all ten actions through the exact same backend endpoints with identical permissions, refactor those endpoints before building another chat interface.