AI Agents Reframing Software: From Fixed Code to Dynamic, On-Demand Systems

A June 2026 arxiv paper (2606.05608) argues that AI agents are restructuring software from 'frozen intent' — pre-written logic encoding anticipated human judgment — into systems that plan and construct behavior on demand, potentially making code no longer the central artifact of software development [1][2]. Kai-Fu Lee and enterprise vendors are amplifying related arguments: multi-agent systems represent the next qualitative step in AI [6], and agent-managed API layers via MCP and competing abstractions are already replacing handwritten integration code [3][5].

If the paper's thesis holds, the fundamental unit of software work shifts from writing and maintaining code to expressing intent that agents interpret and act on dynamically. The near-term practical version is already appearing in API integration products, with active disagreement about which abstraction level — MCP or higher-level 'Skills' — will dominate agent-to-system connections.

A research paper published on arxiv in June 2026 (2606.05608) makes a direct structural argument about software's nature: traditional software is 'frozen intent,' meaning a human anticipated a situation, translated judgment into rules, and shipped fixed code [1]. AI agents, the paper argues, dissolve this structure — software can now plan and construct its own behavior dynamically rather than executing pre-written logic, and code may no longer be the central artifact of software development [2]. The paper has circulated widely in developer and AI commentary circles, with commentators like Rohan Paul positioning it as a near-future transformation rather than speculation [1].

At the infrastructure layer, the MCP (Model Context Protocol) is emerging as a proposed standard for agent-to-API communication. Parloa launched a product called Agent Skills built on MCP, explicitly framed as replacing the brittle glue code developers currently write to handle API authentication, error handling, retries, and data mapping [3]. The pitch is that an MCP-based abstraction layer makes these connections self-managing rather than hand-maintained. Enterprise vendor DataRobot frames the same transition at the product level: from static software products to dynamic AI systems [4]. However, at least one practitioner argues the direction is already moving past MCP toward higher-level 'Skills' abstractions that give enterprises more integration control [5], suggesting MCP's window as the dominant standard may be narrow.

Kai-Fu Lee, founder of Sinovation Ventures, adds a networked-systems dimension. He draws an explicit analogy: a single AI agent today is like a pre-internet personal computer — functional but isolated. Connecting multiple agents enables shared context, task decomposition, and instant coordination, which Lee argues represents the next fundamental step in AI capability, mirroring the historical impact of networking isolated PCs [6]. This multi-agent framing overlaps with but is distinct from the arxiv paper's thesis: both argue for a qualitative change in what software is, but Lee's emphasis is on agent-to-agent coordination rather than agent-to-code replacement.

Product activity is beginning to reflect these arguments. Locofy is positioning itself as an agentic frontend layer between design tools like Figma and code editors [7], one instance of a broader pattern where agents are inserted as active intermediaries in previously code-centric workflows. Whether these product moves validate the paper's stronger claim — that code itself becomes a secondary artifact — or simply add a new tooling layer on top of existing code workflows remains the central unresolved question across this discussion.