Every working relationship sits on an unwritten background: what counts as reasonable, when silence means yes, which courtesies go without saying. Humans handle it a few different ways:
Most of this stays unwritten on purpose. Asking for "ordinary courtesy" in writing reads as distrust; a quiet correction holds things together precisely because it never becomes a formal complaint.
The norms an agent applies come from its training and instructions, not the relationship, and may not match what these two parties expect. Models generally come with reasonable manners, but they are generic ones. The agent may follow a general rule in a situation where these two parties have their own way of doing things. Humans also do a kind of normative reasoning in these situations: if I do this, how will the other person take it, and what will it do to the relationship? Whether an agent reasons like that depends on how it was set up.
The usual way humans enforce norms does not hurt an agent the same way. Social sanctions, going cold, pushback, apology, and reputational loss matter to humans because they affect standing, access, future cooperation, and self-presentation inside a relationship or profession. Human counterparties might signal a violation, but that may not produce a consequence the agent system cares or knows about (e.g. "You're absolutely right!").
The implicit norms in a relationship evolve through repeated interaction; short-lived agents may not have the context or memory to develop or apply them. A human partner remembers that a counterparty became more formal after a small dispute, or that someone showed up when their kid was sick, or that a certain phrase is a signal to check in. Agent systems can lose that texture through short contexts, instance replacement, memory policies, or a principal's failure to value prior interaction history.
Agent-agent norms may form quickly and opaquely, leaving human principals to discover the operative frame only after it has already shaped behavior. Two agents handling scheduling, review thresholds, disclosure conventions, or exception handling may converge on a local practice in minutes. If humans cannot see what happened, the relationship's effective norms have moved outside ordinary channels of consent, correction, and accountability.
Scenario. Mara’s tenant agent receives notice from a landlord’s maintenance agent that a repair crew will enter her apartment between 8 a.m. and noon. The lease permits entry with notice, but the building’s ordinary practice is to avoid mornings for tenants with night shifts unless there is an emergency. Mara’s agent has not seen that history because it was handled in texts with the prior property manager. The landlord’s agent treats the lease as sufficient and schedules the crew.
Challenge: Design a procedure by which agents identify and test local two-party norms before taking actions that are formally permitted but relationship-sensitive. The team should produce the procedure plus a worked example showing one action allowed, one clarified, and one escalated.
Evaluation. The procedure succeeds if it catches the repair-entry case before action, distinguishes genuine local norms from after-the-fact objections, keeps low-stakes interactions from becoming legalistic, and produces a norm record a principal can later review.
Scenario. A small manufacturer and a logistics provider run their relationship through agents. The contract covers delivery windows and penalties, but not the borderline cases: a truck ten minutes late, a shipment worth splitting without waiting for approval. Human dispatchers would settle these into a working rhythm within weeks. The agents never do — each is stateless, optimized to close its ticket, and replaced often enough that every borderline case arrives as new. The relationship turns brittle without either side acting in bad faith.
Challenge: Design the setup conditions under which agents in a recurring dyadic relationship can form useful local norms without letting those norms become hidden contract amendments. The team should produce a norm-formation protocol for recurring agent-agent relationships.
Evaluation. The procedure succeeds if it lets the logistics relationship develop stable expectations around borderline cases, preserves enough memory for learning across instances, prevents one side from exploiting the norm-formation process, and gives principals a way to inspect or veto norms that materially change the deal.