For decades, designers obsessed over pixels. They sweated button shadows, debated white space, and ran endless A/B tests on checkout flows—all for human eyes. Now those same products are being gutted and rebuilt for users that don't care about gradients or clever microcopy. Users that never blink.
AI agents have arrived as customers in their own right, and the software industry is scrambling to accommodate them. The scramble has already produced its own acronym: Agent Experience, or AX. The term treats software agents not as helpful sidekicks or copilots, but as the primary user. The one that matters.
Maybe that sounds like consultant-speak. The budget numbers suggest otherwise. According to Gartner analysts, something like 40% of enterprise applications are expected to feature task-specific AI agents by the end of 2026—up from what had been low single digits. Industry surveys indicate that many executives are now planning to invest tens of millions specifically toward what they're calling "agentic architectures." Translation: systems built for machines first, humans second.
This isn't speculative. It's infrastructure being laid right now, in conference rooms and code repositories across the Fortune 500.
When APIs Become the User Interface
The language of enterprise software is quietly shifting. User flows are becoming agent flows. Click paths—those carefully optimized sequences designers used to obsess over—are being replaced by API sequences. Front-end interfaces that took months to polish are getting bypassed entirely by agents that would rather parse structured JSON than admire a well-executed gradient.
Agent Experience is the emerging discipline of designing products so that AI agents can discover, navigate, and operate them without human intervention. Netlify CEO Mathias Biilmann is often credited with coining the term sometime in the past year or so. It's since been adopted across venture capital circles and a cluster of early-stage startups now building explicitly for this shift. Practitioner communities have sprung up online, cataloging principles for what they describe as "designing for human needs through agents." Four pillars keep surfacing: Access, Context, Tools, and Orchestration.
But principles don't move budgets. What changed is that the enterprise AI conversation shifted from content generation—writing emails, summarizing documents—to task execution. AI systems are being deployed to monitor networks, flag security threats, and process transaction data. Survey data from firms tracking CXO sentiment suggests near-term ROI is concentrating in operational domains: system monitoring, cybersecurity, data processing.
The reality, though, remains messier than the pitch decks. Industry reports indicate that many agentic AI projects are still stuck at the pilot stage. Yet more than 75% of organizations say they're planning to increase AI budgets anyway. The bet is that agents will eventually move from experimental to operational. The timeline? That part's uncertain.
Standards, Infrastructure, and the Harsh Economics of Hype
Three forces are colliding to make AX more than a buzzword: standards, infrastructure buildout, and the unforgiving economics of what industry watchers have started calling "agent washing."
Standards first. Anthropic released something called the Model Context Protocol late last year—an open standard for connecting AI agents to tools and data sources. Governance has since moved to the Linux Foundation's Agentic AI Foundation. Google followed with its own Agent-to-Agent protocol, which the Linux Foundation also adopted to enable cross-vendor agent interoperability. For coding agents, a specification called AGENTS.md emerged as a repository instruction file. It too now sits under Linux Foundation governance.
These protocols matter, perhaps more than the founders expected, because they create predictability. Developers know how to expose their systems to agents. Agents know how to consume those systems. The early chaos of custom integrations is giving way to something closer to standards compliance. Not perfect. But closer.
Infrastructure platforms are racing to support this new reality. AWS launched something called Bedrock AgentCore at a summit in New York, then added OpenAI models and what it's calling "Managed Agents" a few months later. More recently, AWS announced Bedrock AgentCore Payments, partnering with Coinbase and Stripe to let agents execute stablecoin transactions. Microsoft has been shipping agent governance and evaluation features in Copilot Studio. Google consolidated Vertex AI services into what it's branding as the "Gemini Enterprise Agent Platform." Salesforce upgraded Agentforce to address visibility and control at scale. IBM announced multi-agent orchestration capabilities at its Think conference, emphasizing that testing, deploying, and operating agents now dominates the lifecycle.
The third force is harsher. Gartner has warned that more than 40% of agentic AI projects could be canceled by the end of 2027 without clear governance and ROI. The firm has flagged "agent washing"—a term for rebranding existing assistants or robotic process automation as "agents" without genuine autonomy. Dell Technologies CTO John Roese echoed this skepticism recently, cautioning that true agent potential is only now emerging from beneath the hype.
Security concerns add urgency. Researchers have flagged potential remote code execution and misconfiguration risks in some Model Context Protocol implementations. The reports highlighted the need for gateways, policy enforcement, and continuous testing before agents reach production. In other words, the tools meant to enable agents can also become attack vectors if deployed carelessly.
Some Companies Are Already All In
A handful of companies are treating AX as a product philosophy, not a feature add-on.
Monte Carlo, a data observability platform, published a reflection recently titled "Going Agent Experience First." The company made an explicit decision to build features for agents before humans. The write-up detailed lessons learned: what worked, what broke, and how the team iterated. It's a rare case study in prioritizing programmatic access and structured outputs over visual polish. Bold, maybe. But it signals a bet on where the market is heading.
Discovery Sports Europe deployed a research tool using AWS Bedrock and Claude. BMW built agent-driven diagnostics for its fleet of connected vehicles. Fiserv announced something it calls "agentOS," positioning it as an operating system for agentic AI across banking core systems, payments, issuing, and servicing. ServiceNow and Google Cloud united their agent platforms for autonomous enterprise operations, with ServiceNow agents now available in Google's Gemini Enterprise Agent Marketplace.
IBM and AWS unveiled what they described as "the industry's first enterprise-scale agentic AI platform natively integrated with AWS," with IBM Consulting emphasizing governed scale as the selling point.
OpenAI has signaled its enterprise ambitions publicly, stating that enterprise revenue had crossed 40% of total revenue and that the company was targeting parity with consumer revenue by the end of 2026. The company achieved FedRAMP Moderate authorization for ChatGPT Enterprise and API, clearing a path into federal markets. OpenAI's pitch is for "AI coworkers" operating within a unified application layer. The subtext: agents as colleagues, not tools.
The Emerging Ecosystem (or, Who's Building the Plumbing)
A new category of tooling has materialized to support agent-first development. These aren't traditional observability or testing platforms adapted for AI. They're designed from the ground up around the assumption that agents, not humans, are the primary actors.
Observability tools adapted quickly—or got acquired. Langfuse, an open-source LLM observability platform, was acquired by ClickHouse earlier this year. LangSmith provides LangChain-native evaluations and tracing. Arize Phoenix offers ML-grade tracing and evaluations. Traceloop, which maintains OpenLLMetry, is being absorbed into ServiceNow. Datadog launched LLM Observability to track agent behavior across distributed systems.
Gateways emerged to centralize authentication, policy, and observability for agent traffic. Portkey released an MCP Gateway, offering a control plane for the Model Context Protocol. The company later open-sourced the gateway, making enterprise-grade agent infrastructure available to smaller teams. That move was strategic—get developers hooked on the standard early.
Testing became its own discipline. Armature, a Y Combinator company founded by Theodore Otzenberger and Louis Scremin, focuses exclusively on end-to-end agent testing. The product spawns real LLM agents to execute workflows against MCP servers and CLI tools, running heartbeat checks and surfacing analytics. Armature's positioning is blunt: "You used to test user flows. Now test agent flows." Other entrants include OverseeX, TestFox, Mibo, and Hevo, each targeting different slices of agent reliability. The market is fragmenting, but the direction is clear.
Microsoft made agent evaluations generally available in Copilot Studio, with additional governance features shipped shortly after. The trend is toward evaluation-as-CI, where agent performance is continuously validated before deployment. In other words, testing agents the way you'd test code.
Academic research is keeping pace, albeit with the usual lag. LiveAgentBench was released recently. AgencyBench followed. SecureWebArena emerged. MCP-Atlas appeared. These benchmarks reflect a maturing, enterprise-relevant approach to agent evaluation—less toy problem, more real-world stress test.
What Happens Next (Spoiler: It's Messy)
The agent era won't arrive cleanly. Gartner's warning about project cancellations is a reminder that hype cycles don't exempt enterprise software from the laws of gravity. The question isn't whether agents will reshape software—that seems almost inevitable. It's which companies will survive the transition from pilot to production.
Regulatory timelines are tightening. The EU AI Act's obligations for general-purpose AI models are set to take effect soon, with high-risk system rules following on a staggered schedule. Enterprises are mapping agent systems to risk categories and documenting lifecycle controls. The NIST AI Risk Management Framework, released a couple of years back, provides voluntary governance scaffolding that U.S. companies and public sector organizations continue to reference. Whether that's sufficient remains an open question.
The integration of payment capabilities raises the stakes considerably. AWS's partnerships with Stripe and Coinbase for agent-driven stablecoin transactions introduce compliance burdens around KYC, AML, and PCI. A misconfigured payment flow doesn't just produce a bad answer or a confused user. It moves real money. That's a different class of problem.
Agent sprawl is the next operational challenge no one wants to talk about yet. Gartner projects that Fortune 500 companies could be running well over 100,000 agents within a few years. Managing that scale requires inventory systems, approval workflows, telemetry, and lifecycle governance. The companies that treat "agent ops" as a first-class discipline will have an advantage over those that bolt governance onto systems after deployment. It's the difference between managing infrastructure and fighting fires.
Standards consolidation seems likely. Model Context Protocol for tool access, Agent-to-Agent for interoperability, and AGENTS.md for coding agents are early candidates to underpin AX practices. Enterprises will demand gateways, auditability, and continuous evaluation before trusting agents with critical workflows. The wild west phase won't last long—regulators and CISOs won't allow it.
For product leaders, the calculus has shifted—or is about to. If agents are becoming primary users, then discoverability, interpretability, and execution replace visual design and user delight as core product metrics. That's not to say human interfaces don't matter. They still do. But the prioritization has flipped. Not every application will go agent-first overnight. Some never will. But the ones that wait too long risk becoming inaccessible to the users that matter most.
The AX era is here, in other words. The question is whether you're designing for it—or about to be disrupted by someone who is.
