Agent vs Agent

If we're approaching a world where agent-to-agent interactions are the norm and agents routinely represent opposing parties in adversarial economic relationships (insurance vs claimant, insurer vs litigator, buyer vs seller), the moat becomes the agent's ability to outperform the other agent. The rational incentive is to find and exploit weaknesses in the opposing system.

In the case of insurance vs claimant: a claimant's lawyer/agent evaluates injuries, collects medical records, pulls relevant case law and drafts a demand letter; the insurance carrier ingests that demand, assesses liability exposure and generates a counteroffer. Both agents reason over the same case evidence, statutes and policy language (likely some difference in prior settlements though within the same universe). Data and information advantages flatten; separation is whose agent reasons better under adversarial pressure. More precisely: whether one agent can identify systematic weaknesses in the other's reasoning in a shorter time frame.

There is a strategic advantage if your defense agent discovers that the plaintiff's litigation agent systematically underweights certain precedents when presented in a specific sequence (if it finds that opposing agents concede earlier when liability is framed around comparative fault rather than direct negligence). The same logic runs in reverse: plaintiff-side agents will map how carrier agents anchor on initial reserves, where they fold under time pressure, and what framing causes them to lowball or overshoot a claim valuation.

Labs, vertical AI companies and firms deploying the agents face the same calculus: labs will be pressured to harden their models against adversarial probing, vertical companies will build counterparty-specific intelligence as a core capability, and the firms themselves will choose AI vendors partly on how well the vendors' agents perform against a specific set of evals (the awareness and importance of evals will increase and extend beyond just the AI companies).

This moat could concentrate at the model layer, but it will likely depend on how well your agent reasons under adversarial pressure and whether you are learning from every interaction faster than the other side. It also posits an interesting question: does jailbreaking and adversarial probing get repositioned from a safety and alignment problem to a corporate incentive?