Version 3

A June 2026 arxiv paper (2606.05608) argues that AI agents are restructuring software by replacing 'frozen intent' — pre-written logic encoding anticipated human judgment — with systems that plan and construct behavior dynamically, potentially making code no longer the central artifact of software development [2][1]. Rohan Paul, who earlier amplified that thesis, has extended it further: the AI model itself is no longer the product; systems like Codex, Perplexity Computer, and Claude Code are orchestration systems — a model paired with an 'agent harness' that defines how the agent loops and acts [3]. Enterprise vendors and practitioners are building on related premises, with MCP-based abstraction layers replacing handwritten API glue code [4] and agentic intermediaries appearing in design-to-code workflows [8]. Kai-Fu Lee frames the next step as multi-agent networking rather than harness architecture [7].

If neither code nor the base model is the central artifact — but rather the orchestration harness — then competitive advantage in AI products shifts to whoever controls the rules governing how agents loop and act. This would make base models a form of infrastructure and put harness designers in a structurally similar position to operating system vendors in an earlier era. Whether the evidence supports that stronger claim, or whether orchestration is simply a new tooling layer on top of existing code workflows, is still unresolved.

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 over several weeks, though it has not attracted documented critical response from software engineering researchers.

Rohan Paul, who helped amplify the paper's thesis, has since developed a related but distinct claim: the AI model itself is no longer the product. He argues that Codex, Perplexity Computer, and Claude Code are all orchestration systems — each pairs a model with an 'agent harness,' which he defines as the rules governing how the agent loops and acts [3]. On this view, the harness design is the differentiating layer, and model capability is infrastructure. This extends the arxiv paper's argument: not only is code no longer the central artifact, but the base model is not the central value either; the orchestration architecture sitting above both is.

At the infrastructure layer, MCP (Model Context Protocol) is emerging as a proposed standard for agent-to-API communication. Parloa launched 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 [4]. At least one practitioner argues this is already moving past MCP toward higher-level 'Skills' abstractions that give enterprises finer integration control [5], suggesting MCP's window as a dominant standard may be narrow. DataRobot frames the same transition at the product level: from static software products to dynamic AI systems [6].

Kai-Fu Lee adds a networked-systems dimension, arguing that single AI agents are like pre-internet personal computers — functional but isolated — and that connecting multiple agents for shared context and task decomposition represents the next qualitative step in AI capability [7]. 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 [8]. Whether these product moves validate the stronger 'code is secondary' thesis or simply add a new tooling layer on top of existing code workflows remains the central unresolved question across this discussion.