Version 5

A June 2026 arxiv paper (2606.05608) argues AI agents are restructuring software by replacing 'frozen intent' — pre-written logic encoding anticipated human judgment — with systems that plan and construct behavior dynamically [2][1]. Rohan Paul has extended this in two directions: the orchestration harness sitting above the base model is the real product [3], and agentic AI makes any pure-software business structurally vulnerable by enabling on-demand reproduction of software products [8]. Bain & Company is applying this operationally, using vibecoding to generate rough replicas of acquisition targets' software during M&A due diligence to test whether a seller's product has genuine defensibility [7] — a practice now receiving secondary coverage across multiple outlets [9][10].

The Bain practice makes the arxiv thesis concrete at the valuation level: if consultants can quickly replicate the functional shape of a software product to assess its defensibility, then software-as-product has weaker moats than the traditional SaaS model assumes. The question of whether orchestration harnesses or multi-agent networking creates the next defensible layer now has a practical commercial edge.

A research paper published on arxiv in June 2026 (2606.05608) makes a 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, though it has not attracted documented critical response from software engineering researchers.

Rohan Paul, who helped amplify the paper's thesis, has 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 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. Kai-Fu Lee adds a different diagnosis: single agents are like pre-internet personal computers — functional but isolated — and multi-agent networking for shared context and task decomposition is the next qualitative step [4].

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 brittle API glue code with self-healing agent-managed connections [5]. At least one practitioner argues this is already moving past MCP toward higher-level 'Skills' abstractions that give enterprises finer integration control [6], suggesting MCP's window as a dominant standard may be narrow.

The most consequential extension of the thesis is economic. Rohan Paul, reporting on a Financial Times article, describes Bain & Company using vibecoding to generate rough AI-made replicas of acquisition targets' software during M&A due diligence — not as perfect clones, but to reveal whether a product's interface, analytics, automation, or workflow logic is easily reproducible [7]. Separately, former Goldman Sachs executive Raoul Pal argues that agentic AI can reproduce any pure-software product on demand, optimize it, and redeploy it to a better market, making standalone SaaS businesses structurally vulnerable [8]. Together, these items ground the arxiv paper's abstract claim in a concrete valuation question: if the functional shape of a software product can be quickly replicated by consultants or competitors using AI agents, then the defensibility of software-as-product depends less on the code itself and more on data, distribution, or network effects that agents cannot easily reproduce.