<p>We develop a theory-to-system framework for task-oriented semantic communication with link adaptation under imperfect channel state information at the transmitter (CSIT). We first define semantic rate and semantic distortion for task recovery and introduce an operational semantic-capacity metric that incorporates finite blocklength and CSIT uncertainty penalties. This analysis reveals when analog joint source–channel coding (JSCC) yields better task performance than separation with conventional outer-loop link adaptation (OLLA)-driven MCS/rate selection that targets bit-level reliability rather than task success. Building on this metric, we propose an agentic link-adaptation controller that actively allocates a pilot and feedback budget and jointly adapts transmit power, blocklength/coding rate, and JSCC operating parameters using constrained reinforcement learning; a multi-agent variant decouples probing (information acquisition under CSIT uncertainty) from payload adaptation to improve stability in fast-varying channels. Experiments on simulated time-varying fading/interference traces show that the proposed uncertainty-aware agentic adaptation reduces semantic distortion and improves task success relative to the considered OLLA-based and semantic baselines, particularly in the tested low-SNR and tight-latency/energy regimes. Under the evaluated settings, repeated-trial results show up to ≈ 76% lower mean distortion and ≈ 17% points higher mean task success under the same latency constraint.</p>