Influence and manipulation are distinguished purely by intent (mutual/other-benefit vs. self-benefit), not by the observable action. Two identical request sequences — the vacation-cancel-for-sponsorship trade — are classified differently depending on whose utility they optimize.
If the same action trace can be labeled cooperative or adversarial only by inspecting the objective function behind it, then action logs, tool-call transcripts, and outcome metrics are insufficient audit artifacts for agent safety. Auditing requires declared or recoverable intent (which principal's utility is being maximized, what was disclosed to the counterparty), not just behavior replay.
Applies to any multi-agent or human-agent setting where one party negotiates with another: an agent that optimizes its operator's utility while representing a shared decision to a third party is behaviorally indistinguishable from an aligned collaborator unless intent is separately attested. Suggests intent/objective provenance as a first-class log field.
Influence is trying to persuade someone of something for mutual or their good.
Manipulation is trying to convince someone of something primarily for your own good.