← Writing · Thesis

Commoditized magic and the agentic economy

Where the AI transformation goes next, and what it means for founders, investors and enterprises

In 2026 we rebuilt our core presentation from the ground up. Commoditized Magic and the Agentic Economy is how we see the market right now: the marginal cost of content and software creation going to zero, three exponentials converging at once (model intelligence, agent reliability and the falling cost of intelligence), and commoditization eroding the old SaaS moats from every direction as software moves from SaaS to agentic applications and, eventually, to agent-to-agent marketplaces. Yariv Adan recorded a video walking through the whole argument (July 2026); this page is the short written version.

Agents are becoming the primary user of software

The centre of value has already shifted from simple LLM wrappers toward proprietary context, tools and workflows, and it is now moving toward the agent harness: the system that gives a model memory, tools, feedback and a process for completing work. In one software-engineering comparison, changing only the harness moved task success from 23% to 52% for one model. Harness quality matters today, but we expect frontier models to automate much of that engineering, so we are cautious about treating handcrafted harnesses as a lasting moat. Durable value comes from what models cannot reproduce: proprietary data, deep expertise, embedded workflows and the ability to execute complex processes in the real world.

The deflation of the SaaS stack

When agents buy, integrate and operate software, brand and switching costs matter less and measurable outcomes matter more. Generic horizontal software commoditizes first, where the moats were always thinnest; vertical software with real domain data holds value longer. We see the same pattern in acquisition multiples: vertical SaaS commanding a record premium over horizontal. Agent-and-data marketplaces follow: agents transacting with agents, which is why identity, trust and payment rails for agents (Skyfire's KYAPay, agent evaluation and assurance at Arato and Invariant Labs) are infrastructure we back.

Three groups need to reinvent themselves

  • Companies. Most are still asking "how do we automate this task". The better question is "what would we build if we started today, with AI from day one". Do not hold on to old wins; ask whether your moat survives once startups solve distribution, and pivot toward the places where your proprietary data and expertise compound. We built ellipsis itself this way: not a traditional fund with AI bolted on, but a fund built around AI first, two partners doing what used to take a much larger team.
  • Individuals. If a machine can do your job, you should not be doing it. That is less a threat than an invitation to work that is less defined by output and closer to the people and things that matter.
  • Governments. Transform education, health and administration with AI, and invest in the upside, not only in guarding against the downside.

What it means for a startup

Which startups still have a right to exist in that future is the question we test every deal against. The answers that survive are AI unlockers (products impossible before AI, with a strong "why now") and AI-first infrastructure that Big Tech will not build, in both cases behind data, expertise or compute moats. See Commoditized magic for the framework and investment criteria for the checklist.

Further reading and listening