Contested-RF SATCOM reconfiguration is an orchestration problem before it is a modeling problem. The agent that classifies a jamming event is rarely the right agent to choose a new modulation, and neither is the one that reports to the operator. Putting all of that into a single prompt and asking a large model to figure it out works for a demo and fails in operations.
On a recent testbed we split the work across three agents. A sensing agent fuses multi-source RF observations and classifies the disturbance (denial, deception, intermittent, environmental). A deconfliction agent resolves frequency contention against the coalition allocation table and proposes a reconfiguration plan. A reporting agent generates the operator brief in tactical narrative, with a citation back to the underlying observations.
Each agent has its own tool surface, its own eval suite, and its own loop posture. The sensing agent runs autonomously; it has to, given the timescales. The deconfliction agent is human-on-the-loop; the operator sees the proposed plan and can intervene. The reporting agent is human-in-the-loop on outbound messages.
The lesson generalizes. Agentic systems benefit from being narrow at the agent level and composed at the orchestrator level. Wide agents are harder to evaluate, harder to keep current, and harder to defend at acceptance review. Each narrow agent gets an eval baseline and a refusal contract. The orchestrator is the place where the workflow lives, and it stays small enough to read in one sitting.
Mean recovery latency on the testbed dropped roughly an order of magnitude versus the human-only baseline. The number is interesting; the architectural pattern is what we're keeping.