Latest
Original reporting, analysis and practical frameworks across AI products, agents, design, engineering and technology business.
Test the workflow behind the format.
Read the operator label before you ship.
What happens after the connection drops?
Fewer handoffs. Better questions.
When the agent comes to you.
Specs, security and search for AI products.
Evals are the product spec AI teams forget to write
Most AI teams treat evaluation as testing after the build. The teams that ship reliably write evals first, because an eval set is the only spec a model can actually fail.
Prompt injection is an interface problem, not only a model bug
No model reliably separates instructions from data. Agent products have to make that boundary visible, limit risky capability combinations and show where a request came from.
When answers replace clicks, SaaS pages have to be worth citing
Software buyers increasingly ask an AI answer engine before they visit a vendor site. The rules for being cited are less mysterious than the industry suggests, and mostly about clarity.
Reversal, agent-ready design systems and generated apps.
Undo is the feature agent products still lack
Agent interfaces are designed for starting work. Few are designed for stopping it or taking it back, and a single undo button gives false control.
Design systems become agent instruction files
Figma has opened its canvas to agents. Skills files turn a design system from a reference document into instructions an agent reads before it edits anything.
When prompts ship apps, design moves to review
Google I/O 2026 showed agents and prompts producing working apps. The designer's job shifts from drawing the first version to deciding what is acceptable.
Three decisions behind a multi-brand portfolio.
One Company or Two Brands? A Decision Test for Founders
A second identity earns its place when it makes a distinct customer decision easier, and can sustain the work behind that promise.
The Multi-Brand Interface: Share Behavior, Protect Meaning
The best shared layer is often invisible: consistent interaction and accessibility, with a distinct reason to recognize each product.
Cross-Promotion Without Confusing the Reader
A portfolio can share an audience. Each link still needs a clear reason, a visible relationship and room for the reader to say no.
New analysis
Multi-Brand Design Systems: Shared Rules, Distinct Identities
A practical model for sharing product behavior while keeping each brand recognizable.
Shopify Polaris: One API Across Product Surfaces
The same component contract can serve different contexts without forcing a single visual identity.
Decagon Deco: An AI-Ready Design System in Practice
A real case of linking Figma, Storybook and coding agents through explicit component rules.
New today
Figma just turned the design system into a production engine.
Weave can now use existing Figma frames as controlled generative workflow nodes.
Sponsored Agents turn the ad click into a conversation.
OpenAI’s new ad format makes the conversion surface interactive instead of static.
Salesforce is betting the enterprise UI becomes a live interface layer.
AIforce exposes governed data, workflows and permissions to AI interfaces.
AI incident disclosure is becoming part of the product trust layer.
OpenAI’s new reporting framework makes uncertainty, impact and mitigation inspectable.
The AI tool picker is disappearing.
Claude’s unified direction shifts routing work from the user to the system.
September 12, 2026
The dashboard-first era is ending. AI analytics is becoming a conversation.
Analytics is moving from finding a dashboard to asking, inspecting evidence and acting.
GPT‑6 Astra is not your product. The capability boundary is.
Product quality still depends on what the system can do, prove and recover from.
The Agents API changes the build-versus-orchestrate decision.
Agent loops and long-running execution are moving behind APIs.
Agent harness portability is becoming infrastructure.
The abstraction is moving above the model.
Permission UX is becoming the real AI safety interface.
Scope, consent and temporary access become core interaction states.
The new AI trust loop is answer → inspect → challenge.
Evidence provenance is becoming part of the interaction.
Full-duplex voice breaks the old turn-taking model.
Interruption and visible agent state become interface problems.
Stop making every AI feature a chat box.
Persistent state and consequential actions need more than conversation.
A production agent sandbox needs more than isolation.
Agents need scoped credentials, recoverable sessions and observable side effects.
Your average latency is lying to you.
Tail latency is where AI products break.
AI pricing is becoming a product system.
Usage, outcomes and credits need infrastructure built for iteration.
Your next SaaS customer may be an agent.
Products need a machine-usable path alongside the human UI.
More from Interface Report.
- A field guide to agent UX: what the interface should reveal, ask and remember.
- The AI product interface is becoming a control surface, not a chat box.
- Agent Loops Need a Review Loop of Their Own
- Agents Need Component Contracts, Not Screenshot Guesswork
- A Clickable Prototype Is No Longer the Hard Part
- The AI Work System Has Two Ledgers
- One Founder, Three Brands: A Product System in Public
- Persistent Agents Need an Identity That Explains Work
- Rabbit OS3 Turns the Device into an Agent Boundary
- Agent Script Exposes the Line Between Reasoning and Rules
- What makes an AI product feel trustworthy before it feels intelligent.
- Before Agents Automate Work, Map the Exceptions
- Specialized Agents Make Service Policy a UX Surface