Primorum

Primorum — Thesis

The bet behind Primorum.

The next generation of software companies will not be built by larger and larger engineering teams. They will be built by capital allocators with one human signing the cheques and a fleet of AI agents shipping the product, doing the support, running the marketing, and closing the books. We are starting that company.

This is a living document. It captures the bet behind Primorum, why we believe 2026 is the year to make it, what we are doing differently, and the conditions under which we should declare the bet wrong and stop.

1. The Bet

We are betting on three things at once. Each is contested. Together they describe a window that is open right now and that may not stay open for long.

1.1 Foundation models crossed the operator threshold in 2026

For most of LLM history the right framing was “AI as assistant” — a human held the work and the model autocompleted parts of it. That framing is no longer correct for a wide class of software-business tasks. The Claude 4.x, GPT-5.x, and Gemini 3.x families can drive multi-step workflows end-to-end with retries, tool use, and graceful failure handling if the workflow is bounded and the success criteria are observable. Customer support triage, churn interventions, copy for SEO, ad creative iteration, weekly metrics reports, code review on small repos, invoice reconciliation — all solved at the engineering level. Not “demoable.” Solved.

The interesting question is no longer “can the model do it.” The interesting question is “what is the smallest unit of business that can run on top of this capability.” We think it is much smaller than people realize.

1.2 The cost of running an AI-operated business has collapsed

Per-task API cost has dropped by roughly an order of magnitude in 24 months and shows no sign of stopping. A support reply that cost ~$0.30 in 2024 inference now costs ~$0.02. A weekly business review that cost ~$5 of context now costs ~$0.40. This is the difference between AI ops being a luxury for VC-funded startups and being cheaper than the cheapest human anywhere on earth.

When the marginal cost of one more customer interaction is near zero, unit economics on micro-businesses change shape. Niches that were too small for a human team to bother with become large enough for an AI-operated team of one.

1.3 Distribution remains the binding constraint

The first two beliefs are widely shared. The third is where most operators get it wrong. Models are commoditizing. Tooling is commoditizing. The thing that is not commoditizing is getting the product in front of the right person at the right moment.

The bet, stated plainly: foundation models in 2026 are good enough and cheap enough to operate a software business end-to-end, and the surviving operators will be the ones who solve distribution first.

2. The Operating Model

Primorum runs on a deliberate inversion of who does what.

AI is the operator. Not the assistant. The product is run by AI agents from idea to support ticket. Code is written by Claude/Codex; copy by Claude; ads by Claude; support by Claude; weekly reports by Claude. The hand-off point between humans and machines is not “AI drafts, human approves.” It is “AI ships, human spot-checks the result.”

The principal is the only human in the loop. One person, doing only the things that cannot be delegated: capital allocation, legal and brand, kill/greenlight decisions at the gates, and strategy revisions when the world changes. Anything not on this list, the principal does not do.

The moment a principal-as-coder mode appears in the workflow, the model is broken — we are no longer operating an AI-native company, we are operating a one-person agency where the AI helps. That is a different and worse business.

3. Portfolio Game Theory

The unit of strategy at Primorum is not the venture. It is the portfolio.

Most aspiring solo founders pick one idea and bet their time on it. They are running one experiment. Even if the idea is good, variance dominates: founder-market fit, timing, distribution luck, regulation. You are pulling one slot machine handle and hoping.

We are pulling many handles. Cheap ones, fast ones, in parallel. Each spike costs $500–$1,500 end-to-end. The whole architecture — a 100+ idea bank, a 5-dimension scoring framework with distribution double-weighted, a spike-first pipeline, and brutal kill gates — exists to make handle-pulling cheap and informative.

Velocity over thoroughness. Information beats outcome. Every killed spike teaches us something about categories, customers, and the AI moats that hold up under contact. The framework is not a religion; we will tune it based on what the spikes reveal.

4. Operating Rules

  1. AI-native distribution. Platform-native, programmatic SEO/content, AI outbound from public data, or piggyback on an existing audience holder. No human-led sales.
  2. AI is the operator. If a workflow requires a human in the loop to function, we re-scope until it does not, or we kill the idea.
  3. Greenfield only. No reviving old projects, no pivoting dead spikes, no “I have this codebase from 2024 we could use.” The legacy is the cost.
  4. Credible path to $1M. Unit economics × distribution × market size, written down on the page. If we cannot write the path, we do not have a venture; we have a feature.

An idea that breaks any of these does not get scored. It gets killed before scoring.

5. What Could Kill the Bet

We name the failure modes so we recognise them rather than rationalising past them: a capability ceiling, distribution exhaustion, API economics dilution, founder-mode creep, and a portfolio that runs cleanly but never produces a real winner. The kill gates on individual spikes are aggressive; the kill gate on Primorum itself is “we look at the data honestly at each portfolio review and update the bet.”

6. Closing

Primorum is a bet that the right shape for software companies in 2026 is small, AI-operated, distribution-first, and run as a portfolio of cheap experiments by a single principal who never does execution work. We do not need to be right about every belief in this document for the bet to pay off — we need to be right that the combination is more likely to produce a real venture per dollar of capital than any of the alternatives available to one person with $20K.

If the bet is right, it is the cheapest way to start the next generation of software companies. If the bet is wrong, we will know early, kill cleanly, and the worst case is a year of education and a small budget spent.

Either way, we ship the next spike.

Questions, signal, or co-conspiracy — brendan@primorum.ai.