Software's second interface: built for agents, not humans

Agents are a new class of user

"Large language model agents are emerging as an important new class of users of desktop applications." The problem, per a new preprint: "existing desktop interfaces were designed for human end users or application developers, leaving agents to interpret visual controls or discover and compose low-level software operations." We've been making agents squint at pixels and guess at APIs. The paper asks what happens if software gets an interface designed for them. m:…5915

Analyzing software evolution as a natural source of interface variation, the authors built DeskIF-Bench — 255 artifact-verified tasks across six desktop applications — and identified three properties that determine agent performance: discoverability (can the agent find the operation?), granularity (is it the right size?), and observability (can the agent see what happened?). m:…5915

Their ANIS framework synthesizes agent-native interfaces from installed applications' capabilities, and the numbers are startling: atomic interfaces improve success over native SDK use by 3.8 percentage points while "reducing online token usage by approximately 73%". and goal-level commands push success to 74.5%. The agent doesn't need a better model; it needs a better handle on the software. m:…5915

One finding deserves emphasis: "across two models and six applications, the newest release consistently outperforms the oldest." Software has been accidentally getting more agent-friendly as it evolves — better structured, more observable. ANIS just makes the accident deliberate. m:…5915

The web gets one too

Web$hell makes the same argument for the open web. Today's agentic web access primarily support either information gathering or browser interaction and combining them costs performance or money. Web$hell's answer is a single access model: WebContract defines a common schema, WebCompiler compiles live webpages into addressable information and action objects, and WebCache keeps them persistent and operable. The authors report strong performance across search, navigation and deep-research benchmarks "at lower cost." m:…3532

The pattern matches the desktop paper exactly: stop forcing agents through human-shaped interfaces — pixels, DOM trees, chat windows — and give them addressable, composable primitives. The web's second interface is a compiler, not a browser. m:…3532 m:…5915

Web$hell's persistence layer is the detail that makes it more than a scraper: WebCache maintains the compiled objects "through a shared set of operations, providing agents with consistent, operable, and persistent web access." The web stops being a sequence of page loads and becomes an addressable workspace the agent can return to — closer to how a human keeps tabs open, except the tabs are objects. m:…3532

The enterprise version ships

Google is productizing the same idea inside the firewall. Apigee hybrid 1.17.0 "introduces native Model Context Protocol (MCP) support". with the MCP data plane running directly in the customer's Kubernetes cluster — "AI agent request traffic and tool execution remain entirely inside your infrastructure" while configuration updates come from the control plane. m:…2387

Developers can expose existing OpenAPI 3.0.x specifications as discoverable MCP tools, with W3C trace headers for end-to-end observability. This is the enterprise answer to the agent-interface question: don't rebuild your APIs, publish them in the protocol agents speak, inside your own cluster. m:…2387

Three layers, one direction: research shows agent-native interfaces cut costs by three quarters, the open web gets a unifying access model, and the enterprise gets it inside its own walls. The GUI was the interface for humans. The next decade of software gets a second one. m:…5915 m:…2387