Are the AI agents buying and selling in real marketplaces actually different from each other, or basically all the same?

This explores whether agents in real marketplaces differ meaningfully in how they're built, or whether most of them are close variants of the same few designs. The short answer is that the collection has no census of deployed marketplaces. None of these notes counts which models or architectures are actually doing the buying and selling. What the corpus does have is a strong argument for why that missing number matters, plus clues about which forces push toward sameness or toward variety.

The reason to care comes from market design, not engineering. Does agent diversity actually improve market competition? argues that when most agents share the same architecture, they also share the same biases. Those biases then add up across the market instead of cancelling out, and competition gets weaker. The surprise is that diversity only partly fixes this. Mixing in different architectures offsets some of the bias but not all of it. So the question isn't just whether agents are biased. It's what mix of designs is present in the market, which is exactly the thing nobody seems to be measuring.

Whether that variety can exist at all depends on platform structure. Will agentic commerce platforms adopt open or closed models? predicts that agentic commerce will split the way payment networks did. Open systems will let many different agents take part, while closed ones will only accept the platform's own agent. A closed marketplace has close to zero architectural diversity by design. The UK competition regulator's position in Should data portability be infrastructure for agent-driven markets? adds a second gate. If your data can't move with you, you can't switch to a different agent, so lock-in reduces diversity even on nominally open platforms.

The engineering notes point in both directions. Variety tends to show up inside each agent rather than between competing ones. Can small language models handle most agent tasks? argues the economically rational agent mixes models: cheap small models handle the routine work and large models are called only when needed. Infrastructure like Can semantic capability vectors replace manual agent routing? is built on the assumption that very different agents will need to find each other and work together. Other forces push toward sameness. How does test-time scaling work at the agent level? finds that most multi-agent performance comes from how many tokens you spend, not from clever coordination. That suggests performance differences between agents may come down to budget more than design. And Why do accuracy-only benchmarks hide the true cost of agents? shows that leaderboards reward everyone for climbing the same accuracy hill, which tends to make designs converge.

The takeaway: 'architectural diversity' could mean two different things. One is variety inside a single agent's stack. The other is variety across the population of agents competing in a market. The research mostly encourages the first, while the market argument depends on the second. A marketplace where every seller's agent is a well-designed mix of small and large models, all from the same vendor, could still be the uniform population Bichler warns about. That's also why Should AI agents serve one user or balance many interests? argues that alignment has to look at outcomes across the whole market, not just loyalty between one agent and its user.