Asymmetric Federated Agentic Intelligence

Enterprise deployments of autonomous language-model agents are moving from single-tenant pilots to workflows that cross institutional boundaries, where each participant runs a different model, exposes a different private tool registry, operates under a different privacy budget, and holds a different level of decision authority. Classical federated learning assumes a shared parameter space and a stationary loss, while multi-agent systems research typically assumes a common environment and symmetric authority, so neither framework describes this regime. We formalize Asymmetric Federated Agentic Intelligence (AFAI): a setting in which structurally unequal institutional agents collaborate by exchanging sanitized cognitive artifacts instead of parameters. We contribute (i) a taxonomy of four orthogonal enterprise asymmetries and a septuple specification of an asymmetric enterprise agent; (ii) differential trajectory privacy, which transfers record-level differential privacy to abstracted reasoning traces, with closure under post-processing and sublinear multi-round accounting; (iii) asymmetric strategy distillation, a local objective whose stationarity gap with respect to the local task is bounded by an explicit distillation-bias term that yields a capacity-dependent regularization rule; (iv) a capability-brokering rule defined only over privacy-dominant and authority-compatible partner sets, with refusal as a first-class outcome; and (v) an evaluation blueprint of six computable metrics, three challenge environments, and a reporting protocol aligned with current agentic-benchmark validity practice. AFAI is a framework and measurement proposal rather than an empirical study, and Section 6 states the resulting limitations explicitly.