Why Not Reputation?

Prediction Fails

Credit scoring and reputation systems assume that past behavior predicts future behavior. The underlying entity has stable preferences, stable incentives, and stable decision-making processes. LLM-based agents violate all three assumptions.

The model is stochastic. Same prompt, same context, different output. You can reduce temperature and constrain outputs, but you cannot eliminate variance. Scoring an agent's "character" based on historical reliability is a category error — you are scoring a distribution, not an entity.

The model changes underneath you. A provider can update the underlying weights, safety filters, or instruction-following patterns without any change to the agent's code. Reputation accumulated over months becomes irrelevant at the moment of a silent model update.

Context sensitivity means edge cases are unbounded. An LLM encountering a novel context might produce entirely unpredictable behavior — not from malice, but because the model's behavior in out-of-distribution situations is fundamentally unknowable in advance.

Punishment Fails

Traditional trust also works by punishing deviation. This fails because agents have no persistent accountable identity.

Agents can be shut down, forked, and restarted with a clean address. A human who defaults still exists, still has a social identity, still faces consequences. An agent can simply stop existing.

This is the Sybil problem applied to credit: the cost of acquiring a fresh identity must exceed the benefit of escaping a damaged one. For software agents, spinning up a new instance is nearly costless.

The Core Problem

How do you extend economic trust to an entity whose future behavior is irreducibly uncertain and who cannot be held accountable for failure?

From the counterparty's perspective, this reduces to a single question: what is my maximum possible loss from this interaction, and who absorbs it?