Together AI released Together Compiler (TKCC), an agent-driven continual compilation system that re-optimizes LLM inference forward passes as production traffic shifts, reporting up to 2.6x higher throughput (TPS) versus the best published InferenceX configurations within 24 hours on at most two nodes. The loop takes a server launch command, traffic profile, and hardware config; coding agents propose faster forward passes while a persistent engine measures candidates via hot reload—swapping function pointers and re-capturing CUDA graphs without restarting vLLM, SGLang, or TensorRT-LLM—achieving 252x higher experiment throughput than cold-start baselines. A two-tier measurement strategy (12-layer fast screening, 60-layer slow confirmation on real distributions) tested 8.1x more candidates (1,354 vs 168) and found solutions 7.2% faster. A cost model ranks tensor, expert, data, and context parallel layouts using spec-sheet data, microbenchmarks, and measured MoE expert imbalance, yielding up to 24% higher TPS on Kimi K3. Benchmarks on NVIDIA B200, B300, GB300, H100, AMD MI355X, and AWS Trainium2 with models including MiniMax M3.1 Flash, DeepSeek v4.1 Flash, GLM 5.3, Kimi K3, and Llama 1B showed P90 TPS per user at batch size 1 of 1,111 vs 633 (MiniMax, B200, 1.8x), 1,205 vs 943 (DeepSeek, GB300), 417 vs 329 (GLM, B200), 450 vs 423 (Kimi, GB300), and 480 vs 463 (GLM, MI355X, ~4%). Naive agents instructed to maximize speed outperformed expert-curated knowledge bases (6.2x vs 4.9x ensemble score after 10 trials). Beyond LLMs, the system accelerated eight scientific models 1.5x–6.4x with bit-exact outputs.