From UX to AX: What Design Looks Like When AI Becomes Your Co-Worker

$800 Billion Disappeared from Software Stocks in Two Weeks. The Interfaces Those Companies Sell Are the Reason.

In the first two weeks of February 2026, the S&P 500 Software Index dropped 13% in five trading sessions. Salesforce and Adobe fell more than 25% since the start of the year. Atlassian cratered. Forrester declared "SaaS as we know it is dead."

The catalyst was not a recession or earnings miss. It was Anthropic's release of agentic capabilities on February 3 ("Black Tuesday for Software"), which the market interpreted as a direct threat to the per-seat SaaS model. The logic was simple: if AI agents can operate software on behalf of users, the value of the interface itself collapses.

This is not just a stock market story. It is a design story. The market is pricing in a fundamental truth that designers and product leaders need to hear: the interface paradigm we have built careers around is being restructured from the ground up.

We are moving from an era where humans operate software to an era where AI agents operate software on behalf of humans. The interface is no longer the product. The interface is the trust layer between you and the agents doing the work.

Some practitioners are calling this shift AX: Agentic Experience.

What Is AX? The Shift from HCI to Human-Agent Orchestration

For 40 years, interface design has been built on a single assumption: the human directly operates the software. Click a button. Fill a form. Navigate a menu. Scan a dashboard. This is Human-Computer Interaction (HCI), the foundation of everything from Xerox PARC to the iPhone.

AX (Agentic Experience) breaks that assumption.

In AX, a built-in agent continuously supports the user across the journey, not just in isolated features. Instead of the user doing all the planning and navigating, the agent shares the cognitive and logistical load: summarizing, planning, and executing steps within the same experience. As UXmatters puts it: "Good design turns autonomy into partnership, while bad design turns autonomy into risk."

Consultancies working on agentic AI describe the shift as moving from designing static flows for abstract personas to designing adaptive systems for individuals in real time. The user no longer navigates to search for information. They entrust a complex mission to an AI.

This is what NNGroup's State of UX 2026 report describes: "In many products, people are spending less time navigating the UI and more time delegating to a layer sitting on top of it."

The Four Capabilities of AX

One influential framing breaks agentic systems into four core capabilities. UX has to surface and constrain each of these in ways that feel legible and safe to humans.

For designers, this shifts focus from linear task flows to flexible loops where humans and agents hand work back and forth across multiple iterations. You are no longer designing a path through screens. You are designing a relationship between a human and an autonomous system.

The Six Patterns Replacing Every Button and Form

The design community has converged on six patterns that follow the functional lifecycle of an agentic interaction. These were documented across Smashing Magazine, UXmatters, and UX Magazine independently and reached consensus.

The Autonomy Dial Is the Breakthrough

The most transformative pattern is the Autonomy Dial. UXmatters describes it as "progressive authorization" where users define boundaries on a spectrum:

• Full manual: User does everything, agent suggests
• Supervised autonomy: Agent acts, user approves
• Bounded autonomy: Agent acts within defined boundaries
• Full autonomy: Agent acts independently within broad goals

This makes users "active participants in defining the relationship" rather than passive recipients of AI decisions. It mirrors what product teams already do with junior colleagues: delegate work, review drafts, approve or reject suggestions, and occasionally step in to handle edge cases directly.

A New Anti-Pattern: Agentic Sludge

Smashing Magazine also coined the term "Agentic Sludge" to describe a new dark pattern: removing friction to a fault, making it too easy for a user to agree to an action that benefits the business rather than their own interests. When an agent can execute actions with a single click, the design ethics around consent become more critical, not less.

Intent-First Interfaces: The Death of the Menu

A consistent pattern across the literature is the move from feature-first UIs (menus, filters, wizards) to intent canvases where the primary interaction is "state your goal and constraints."

Instead of choosing from many tools, users increasingly specify outcomes: "analyze this research set," "prepare next week's outreach," "triage this incident." The agent composes the workflow.

In UI terms, this shows up as dedicated areas for:

1. Goal definition: What the user wants to accomplish
2. Constraints and policy selection: Boundaries, budgets, permissions
3. Shared context: Files, channels, tools the agent can access
4. Live plan and progress timeline: What the agent is doing right now

CIO.com describes this as the end of dashboards: "What once revolved around dashboards, reports, and retrospective analysis is rapidly evolving into contextual, autonomous, and action-oriented intelligence systems."

The dashboard is not dead. It is evolving. Static reports become conversational querying. Manual chart-building becomes agents that surface insights proactively. The interface still exists, but its job changes from "show me everything" to "show me what matters right now."

Generative UI: Interfaces That Build Themselves

The most radical shift is generative UI, where the interface itself is created dynamically by the AI agent. Instead of one static interface for millions of users, the system evaluates real-time intent and assembles relevant components on the fly.

Google Research has rolled out two experiments in the Gemini app: "dynamic view" and "visual layout," where Gemini designs and codes a fully customized interactive response for each prompt.

There are three tiers of generative UI, per CopilotKit:

Vercel AI SDK 6 made this developer-ready with 25+ purpose-built React components called AI Elements. Shopify's MCP-UI renders interactive product cards, cart flows, and booking forms directly inside agent conversations.

NNGroup warns about the trade-off: AI hallucinations "would result in the same problems in a generative UI," and constantly changing layouts could create "frustration" as users must relearn interfaces repeatedly. Designers must "balance the gains from a completely customized experience with the losses incurred by the lack of UI consistency and predictability."

Every Major Product Is Already Redesigning

This is not theory. Every major software company is rebuilding their interfaces around agents.

Some of the most dramatic transformations:

Salesforce rebuilt their entire platform around Agentforce, with an Agent Canvas (no-code visual editor), Agent Script (a new markup language for agent behavior), and a real-time simulation panel where users test agent reasoning live. Their stated thesis: "AI agents will become the new UI, and apps take a backseat."

Microsoft put agents on the Windows taskbar. When AI agents run longer tasks, they now appear at the OS level showing at-a-glance status and chain-of-thought logic. This is a new OS-level UI primitive.

Notion 3.0 launched autonomous agents that work for up to 20 minutes independently, performing multi-step tasks across hundreds of pages. Users can choose between GPT-5.2, Claude Opus 4.5, and Gemini 3 to power their agents. The model selector became a first-class UI element.

Slack transformed from a messaging app into an agent workspace. Agentforce agents appear in Slack "just like they would with a teammate." The agent sits in the sidebar like any other contact.

Apple is preparing its most dramatic Siri overhaul: a project codenamed Campos with a fresh architecture and interface designed from the ground up for the chatbot era, expected with iOS 27 at WWDC 2026.

The Four Protocols Standardizing Agentic Interfaces

For the first time, the industry is converging on open protocols that define how agents generate, communicate through, and control interfaces.

The most interesting for designers is Google's A2UI (Agent-to-User Interface), released December 2025. It is a declarative data format (not executable code) where agents describe interfaces using a flat list of components with ID references. Your app maintains a "catalog" of trusted, pre-approved UI components, and the agent can only request components from that catalog. This is security by design: the agent cannot render arbitrary code, only select from what you have approved.

CopilotKit's AG-UI complements this by standardizing the real-time communication layer. Instead of waiting for a final answer, the agent sends approximately 16 standard event types describing what it is doing as it works.

For product teams, the implication is clear: if you are building anything that involves AI agents interacting with users, you need to understand this stack. It is becoming the foundation, the same way REST APIs became the foundation for web services.

Trust Is THE Design Challenge of 2026

Every source we reviewed circled back to the same conclusion: trust is the central design problem.

NNGroup's State of UX 2026 states it directly: "In 2026, trust will be a major design problem for AI experiences. This challenge will only grow as more and more AI agents are rolled out, often before they are ready."

CMSWire goes further: "Trust is the new benchmark for AI, and UX owns the outcome."

A CSCW 2025 study provided empirical evidence: higher transparency significantly improved user trust, satisfaction, and willingness to use AI design agents, using the TOAST scale (Trust of Automated Systems Test) across three levels of transparency.

The Uncanny Valley of Mind

Here is the finding that should change how every product team designs their AI: a Frontiers in Psychology systematic review (2025) confirmed the uncanny valley extends into text-based and conversational AI. "Uncanny-like" bots that are "almost, but not quite" human were consistently rated lowest on trust, comfort, and willingness to interact.

Research on AI service recovery found that the positive influence of anthropomorphism is entirely contingent on flawless execution. Minor glitches in highly human-like agents negate all benefits by violating expectations of authentic interaction.

The practical takeaway: the more human you make your AI seem, the more catastrophically it fails when it makes mistakes. Users actually felt more secure with a mechanized, clearly non-human chatbot.

The right level of anthropomorphism is: clearly non-human but communicatively intelligent. Functional, not performative.

The Co-Worker vs Tool Debate

Should AI agents be designed as teammates or instruments? The industry has not reached consensus.

The co-worker camp: Salesforce and Slack design agents that appear alongside human contacts with names, avatars, and conversational interfaces. Notion positions its agents as autonomous workers. GitHub calls Copilot a "development partner."

The instrument camp: IDC published a direct counterpoint: "The popular narrative of AI as a 'co-worker' oversells its role and misunderstands its limits. AI systems are not peers; they are instruments: programmable, bounded, and entirely dependent on human judgment."

The pragmatic middle ground, and the one the research supports, treats agents as capable instruments with personality. Named identity (for natural interaction). Emotional intelligence (for appropriate responses). Transparent about being AI (for trust). Proactive but deferential (for productive collaboration).

This is exactly the approach we took with Vector, our AI sales agent. Vector has a named identity, 11 negotiation patterns with emotional intelligence, and proactive behavior triggered by exit intent and page context. But it is transparent about being AI, and it knows when to escalate to our human founders. The co-worker framing works because it sets the right expectations without over-promising.

Relationship UX: The Long Game

One of the most underexplored angles in the AX conversation is what some practitioners call "Relationship UX." Instead of designing for single-session interactions, agentic systems create ongoing relationships between humans and agents.

Design implications include:

• Surfaces for long-term memory: Interfaces that expose projects, preferences, and prior decisions the agent is using, with options to correct or forget them
• Relationship rituals: Recurring reviews, standups, or retrospectives between user and agent (a weekly pipeline review, a monthly campaign analysis)
• Persona-driven multi-pattern GUIs: Successful AI products will pick a strong persona (e.g., "ops lead," "SDR teammate") and use multiple agentic patterns inside one GUI to serve that role well

This nudges UX work closer to service design: thinking about long-term engagement, trust, and growth of capability between human and agent over months, not just first-run flows.

What This Means for Designers

The designer role is splitting. According to Designative's Design Specialization Impact Matrix, the automation risk varies dramatically by specialization.

Figma's 2025 AI Report (2,500 users across seven countries) reveals the tension: 78% of designers say AI makes them more efficient, but only 47% say it makes them better at their jobs. Only 32% trust AI output. Three times more designers feel the job market has declined compared to the previous year.

Yet 85% say learning to work with AI will be essential to future success.

Jakob Nielsen describes the future of design work as resembling service design more than UI design. Rather than designing screens, future designers will focus on "policy surfaces" (permissions, cost limits, ethical boundaries), "confidence conveyors" (showing provenance, uncertainty, rollback options), and "system temperament" (agent behavior patterns). He calls this "agent choreography" and "experience governance."

The Conversation Design Institute identifies the key emerging role as AI Orchestration Designer: designing how multiple AI agents, tools, and systems coordinate to solve complex tasks. This is "the traffic controller for AI," mapping orchestration patterns, handoffs, escalation, and tool use into usable experiences.

The Skills That Matter Now

The new toolkit blends information architecture, conversation design, operations thinking, and behavioral economics. Key skills include:

• Systems thinking: Designing rule systems, not individual screens
• Interaction choreography: Multi-step, mixed-initiative interactions instead of linear flows
• Policy and safety UX: Co-designing with legal, risk, and ops to express guardrails in understandable ways
• Explanation copywriting: Treating explanation copy with the same care as UI copy
• Multimodal prototyping: Designing across chat, voice, screens, and automated workflows

The Reality Check

The shift is real, but the hype outpaces adoption. Deloitte's State of AI in the Enterprise 2026 provides hard numbers:

Critics describe AI agents as "junior staffers who work quickly, confidently and often incorrectly" (Orlando Sentinel). The Klarna headline about "replacing Salesforce with AI" turned out to be more nuanced. CX Today reported they actually switched to alternative SaaS tools like Deel, consolidating their stack rather than replacing it with pure AI. Companies report new inefficiencies from agentic AI: duplicated work, increased oversight burdens, and time spent correcting errors.

NNGroup warns that generative UI at scale may be decades away due to consistency and processing power constraints. The prediction from Futurum Group: "By the end of 2026, the conversation around AI agents will begin to mature, the hype will cool, and executives will talk less about autonomy and more about supervision and co-piloting."

None of this invalidates the shift. It just means the transition will be messy, incremental, and full of false starts. The 11% that have agents in production today are building the institutional knowledge that the other 89% will need.

Where This Is Going

The trajectory points in one direction: less screen time, more delegation.

OpenAI is developing a screenless, voice-first hardware device with Jony Ive for Fall 2026. Lenovo and Motorola launched Qira at CES 2026, a "Personal Ambient Intelligence" system that follows users across their entire device ecosystem.

Jakob Nielsen predicts generative UI becomes the dominant paradigm, UX becomes the primary business moat as raw AI capabilities commoditize, and 2026 will be remembered as the year the infrastructure of the AI era was laid.

The opportunity for designers, product leaders, and agencies is enormous. No major design system (Material Design, Apple HIG, Fluent, Ant Design) has published a dedicated "agentic AI" section as of February 2026. Whoever publishes the definitive agentic design guidelines will shape the next decade of interfaces, the same way Material Design defined mobile interaction.

The interface is not dying. It is becoming something new: a trust layer, an orchestration surface, a relationship medium between humans and the agents that work alongside them.

For a studio like ours, this is not abstract theory. It is what we build every day. From Vector's 11 negotiation patterns to Hive's multi-agent co-worker platform, we design the trust, transparency, and delegation surfaces that make agentic AI actually useful in production.

The question is not whether AX will replace traditional UX. It already is. The question is whether you will lead the transition or follow it.