diff --git a/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c16-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c16-mtp.yaml
new file mode 100644
index 000000000..a3f8bb681
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c16-mtp.yaml
@@ -0,0 +1,175 @@
+# Experimental; retain only measured improvements over the current Pareto frontier.
+schema: 2
+name: qwen35-b200-fp8-agentx-disagg-1p1d-p-tp4-d-tp4-hicache-c16-mtp
+model:
+  path: qwen3.5-fp8
+  container: lmsysorg/sglang:nightly-dev-cu13-20260918-20518d85
+  precision: fp8
+engine: sglang
+dynamo:
+  install: true
+  source:
+    rev: 805a77f053d807b0d8def5d27f674a6df0ed839e
+resources:
+  gpu_type: b200
+  gpus_per_node: 8
+slurm:
+  time_limit: '4:00:00'
+frontend:
+  type: dynamo
+  nginx_container: nginx-sqsh
+  nginx_session_affinity: true
+  nginx_session_affinity_header: X-Dynamo-Session-ID
+  enable_multiple_frontends: true
+  num_additional_frontends: 1
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: '1'
+    PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+  args:
+    router-mode: kv
+    router-session-affinity-ttl-secs: '3600'
+roles:
+  prefill:
+    nodes: 1
+    workers: 1
+    gpus: 4
+    env:
+      PYTHONUNBUFFERED: '1'
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+      SGLANG_ENABLE_SPEC_V2: '1'
+      TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+      SGLANG_TIMEOUT_KEEP_ALIVE: '1800'
+      NCCL_NVLS_ENABLE: '1'
+      SGL_ENABLE_JIT_DEEPGEMM: 'false'
+      SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+      PYTHONNOUSERSITE: '1'
+      MC_INTRANODE_NVLINK: '1'
+      SGLANG_MOONCAKE_CUSTOM_MEM_POOL: INTRA_NODE_NVLINK
+    args:
+      served-model-name: Qwen/Qwen3.5-397B-A17B-FP8
+      trust-remote-code: true
+      tensor-parallel-size: 4
+      expert-parallel-size: 1
+      data-parallel-size: 1
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+      mamba-ssm-dtype: bfloat16
+      attention-backend: trtllm_mha
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      page-size: 64
+      mem-fraction-static: 0.8
+      context-length: 262144
+      chunked-prefill-size: 16384
+      max-prefill-tokens: 16384
+      max-running-requests: 32
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      enable-metrics: true
+      enable-cache-report: true
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      disable-cuda-graph: true
+      enable-symm-mem: true
+      scheduler-recv-interval: 10
+      stream-interval: 50
+      mamba-radix-cache-strategy: extra_buffer
+      mamba-track-interval: 2048
+      enable-hierarchical-cache: true
+      hicache-size: 104
+      hicache-io-backend: kernel
+      hicache-mem-layout: page_first
+      hicache-write-policy: write_through_selective
+  decode:
+    nodes: colocate
+    workers: 1
+    gpus: 4
+    env:
+      PYTHONUNBUFFERED: '1'
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+      SGLANG_ENABLE_SPEC_V2: '1'
+      TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+      SGLANG_TIMEOUT_KEEP_ALIVE: '1800'
+      NCCL_NVLS_ENABLE: '1'
+      SGL_ENABLE_JIT_DEEPGEMM: 'false'
+      SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+      PYTHONNOUSERSITE: '1'
+      MC_INTRANODE_NVLINK: '1'
+      SGLANG_MOONCAKE_CUSTOM_MEM_POOL: INTRA_NODE_NVLINK
+    args:
+      served-model-name: Qwen/Qwen3.5-397B-A17B-FP8
+      trust-remote-code: true
+      tensor-parallel-size: 4
+      expert-parallel-size: 1
+      data-parallel-size: 1
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+      mamba-ssm-dtype: bfloat16
+      attention-backend: trtllm_mha
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      page-size: 64
+      mem-fraction-static: 0.8
+      context-length: 262144
+      chunked-prefill-size: 16384
+      max-prefill-tokens: 16384
+      max-running-requests: 32
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      enable-metrics: true
+      enable-cache-report: true
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      disable-radix-cache: true
+      speculative-algorithm: NEXTN
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+      cuda-graph-max-bs-decode: 16
+      stream-interval: 50
+      enable-symm-mem: true
+      scheduler-recv-interval: 10
+      mamba-radix-cache-strategy: no_buffer
+      mamba-track-interval: 128
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+sbatch_directives:
+  mem: '0'
+srun_options:
+  mem: '0'
+  container-remap-root: ''
+telemetry:
+  enabled: true
+  collect_interval_ms: 1000
+  storage_subdir: power
+  required: true
+  startup_timeout_seconds: 120
+  request_timeout_seconds: 2
+  collector_join_timeout_seconds: 12
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    port: 9401
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  concurrencies:
+  - 16
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    SRT_MEASUREMENT_WINDOW_BENCHMARK_TYPE: custom
+    SRT_MEASUREMENT_WINDOW_CONCURRENCIES: '16'
+    SRT_MEASUREMENT_WINDOW_RESULT_ROOT: /logs
+    PORT: '8000'
+    IS_MULTINODE: 'true'
+    EVAL_FRAMEWORK: lm-eval
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126_256k
+    AIPERF_MAX_CONTEXT_LENGTH: '262144'
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: 'true'
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0'
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'sglang:'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c24-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c24-mtp.yaml
new file mode 100644
index 000000000..a9af6f125
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c24-mtp.yaml
@@ -0,0 +1,175 @@
+# Experimental; retain only measured improvements over the current Pareto frontier.
+schema: 2
+name: qwen35-b200-fp8-agentx-disagg-1p1d-p-tp4-d-tp4-hicache-c24-mtp
+model:
+  path: qwen3.5-fp8
+  container: lmsysorg/sglang:nightly-dev-cu13-20260918-20518d85
+  precision: fp8
+engine: sglang
+dynamo:
+  install: true
+  source:
+    rev: 805a77f053d807b0d8def5d27f674a6df0ed839e
+resources:
+  gpu_type: b200
+  gpus_per_node: 8
+slurm:
+  time_limit: '4:00:00'
+frontend:
+  type: dynamo
+  nginx_container: nginx-sqsh
+  nginx_session_affinity: true
+  nginx_session_affinity_header: X-Dynamo-Session-ID
+  enable_multiple_frontends: true
+  num_additional_frontends: 1
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: '1'
+    PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+  args:
+    router-mode: kv
+    router-session-affinity-ttl-secs: '3600'
+roles:
+  prefill:
+    nodes: 1
+    workers: 1
+    gpus: 4
+    env:
+      PYTHONUNBUFFERED: '1'
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+      SGLANG_ENABLE_SPEC_V2: '1'
+      TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+      SGLANG_TIMEOUT_KEEP_ALIVE: '1800'
+      NCCL_NVLS_ENABLE: '1'
+      SGL_ENABLE_JIT_DEEPGEMM: 'false'
+      SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+      PYTHONNOUSERSITE: '1'
+      MC_INTRANODE_NVLINK: '1'
+      SGLANG_MOONCAKE_CUSTOM_MEM_POOL: INTRA_NODE_NVLINK
+    args:
+      served-model-name: Qwen/Qwen3.5-397B-A17B-FP8
+      trust-remote-code: true
+      tensor-parallel-size: 4
+      expert-parallel-size: 1
+      data-parallel-size: 1
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+      mamba-ssm-dtype: bfloat16
+      attention-backend: trtllm_mha
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      page-size: 64
+      mem-fraction-static: 0.8
+      context-length: 262144
+      chunked-prefill-size: 16384
+      max-prefill-tokens: 16384
+      max-running-requests: 48
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      enable-metrics: true
+      enable-cache-report: true
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      disable-cuda-graph: true
+      enable-symm-mem: true
+      scheduler-recv-interval: 10
+      stream-interval: 50
+      mamba-radix-cache-strategy: extra_buffer
+      mamba-track-interval: 2048
+      enable-hierarchical-cache: true
+      hicache-size: 104
+      hicache-io-backend: kernel
+      hicache-mem-layout: page_first
+      hicache-write-policy: write_through_selective
+  decode:
+    nodes: colocate
+    workers: 1
+    gpus: 4
+    env:
+      PYTHONUNBUFFERED: '1'
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+      SGLANG_ENABLE_SPEC_V2: '1'
+      TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+      SGLANG_TIMEOUT_KEEP_ALIVE: '1800'
+      NCCL_NVLS_ENABLE: '1'
+      SGL_ENABLE_JIT_DEEPGEMM: 'false'
+      SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+      PYTHONNOUSERSITE: '1'
+      MC_INTRANODE_NVLINK: '1'
+      SGLANG_MOONCAKE_CUSTOM_MEM_POOL: INTRA_NODE_NVLINK
+    args:
+      served-model-name: Qwen/Qwen3.5-397B-A17B-FP8
+      trust-remote-code: true
+      tensor-parallel-size: 4
+      expert-parallel-size: 1
+      data-parallel-size: 1
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+      mamba-ssm-dtype: bfloat16
+      attention-backend: trtllm_mha
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      page-size: 64
+      mem-fraction-static: 0.8
+      context-length: 262144
+      chunked-prefill-size: 16384
+      max-prefill-tokens: 16384
+      max-running-requests: 48
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      enable-metrics: true
+      enable-cache-report: true
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      disable-radix-cache: true
+      speculative-algorithm: NEXTN
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+      cuda-graph-max-bs-decode: 24
+      stream-interval: 50
+      enable-symm-mem: true
+      scheduler-recv-interval: 10
+      mamba-radix-cache-strategy: no_buffer
+      mamba-track-interval: 128
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+sbatch_directives:
+  mem: '0'
+srun_options:
+  mem: '0'
+  container-remap-root: ''
+telemetry:
+  enabled: true
+  collect_interval_ms: 1000
+  storage_subdir: power
+  required: true
+  startup_timeout_seconds: 120
+  request_timeout_seconds: 2
+  collector_join_timeout_seconds: 12
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    port: 9401
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  concurrencies:
+  - 24
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    SRT_MEASUREMENT_WINDOW_BENCHMARK_TYPE: custom
+    SRT_MEASUREMENT_WINDOW_CONCURRENCIES: '24'
+    SRT_MEASUREMENT_WINDOW_RESULT_ROOT: /logs
+    PORT: '8000'
+    IS_MULTINODE: 'true'
+    EVAL_FRAMEWORK: lm-eval
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126_256k
+    AIPERF_MAX_CONTEXT_LENGTH: '262144'
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: 'true'
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0'
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'sglang:'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c32-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c32-mtp.yaml
new file mode 100644
index 000000000..5dab4917a
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c32-mtp.yaml
@@ -0,0 +1,175 @@
+# Fast replay improves the published frontier; canonical sweep qualification is pending.
+schema: 2
+name: qwen35-b200-fp8-agentx-disagg-1p1d-p-tp4-d-tp4-hicache-c32-mtp
+model:
+  path: qwen3.5-fp8
+  container: lmsysorg/sglang:nightly-dev-cu13-20260918-20518d85
+  precision: fp8
+engine: sglang
+dynamo:
+  install: true
+  source:
+    rev: 805a77f053d807b0d8def5d27f674a6df0ed839e
+resources:
+  gpu_type: b200
+  gpus_per_node: 8
+slurm:
+  time_limit: '4:00:00'
+frontend:
+  type: dynamo
+  nginx_container: nginx-sqsh
+  nginx_session_affinity: true
+  nginx_session_affinity_header: X-Dynamo-Session-ID
+  enable_multiple_frontends: true
+  num_additional_frontends: 1
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: '1'
+    PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+  args:
+    router-mode: kv
+    router-session-affinity-ttl-secs: '3600'
+roles:
+  prefill:
+    nodes: 1
+    workers: 1
+    gpus: 4
+    env:
+      PYTHONUNBUFFERED: '1'
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+      SGLANG_ENABLE_SPEC_V2: '1'
+      TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+      SGLANG_TIMEOUT_KEEP_ALIVE: '1800'
+      NCCL_NVLS_ENABLE: '1'
+      SGL_ENABLE_JIT_DEEPGEMM: 'false'
+      SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+      PYTHONNOUSERSITE: '1'
+      MC_INTRANODE_NVLINK: '1'
+      SGLANG_MOONCAKE_CUSTOM_MEM_POOL: INTRA_NODE_NVLINK
+    args:
+      served-model-name: Qwen/Qwen3.5-397B-A17B-FP8
+      trust-remote-code: true
+      tensor-parallel-size: 4
+      expert-parallel-size: 1
+      data-parallel-size: 1
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+      mamba-ssm-dtype: bfloat16
+      attention-backend: trtllm_mha
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      page-size: 64
+      mem-fraction-static: 0.8
+      context-length: 262144
+      chunked-prefill-size: 16384
+      max-prefill-tokens: 16384
+      max-running-requests: 64
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      enable-metrics: true
+      enable-cache-report: true
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      disable-cuda-graph: true
+      enable-symm-mem: true
+      scheduler-recv-interval: 10
+      stream-interval: 50
+      mamba-radix-cache-strategy: extra_buffer
+      mamba-track-interval: 2048
+      enable-hierarchical-cache: true
+      hicache-size: 104
+      hicache-io-backend: kernel
+      hicache-mem-layout: page_first
+      hicache-write-policy: write_through_selective
+  decode:
+    nodes: colocate
+    workers: 1
+    gpus: 4
+    env:
+      PYTHONUNBUFFERED: '1'
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+      SGLANG_ENABLE_SPEC_V2: '1'
+      TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+      SGLANG_TIMEOUT_KEEP_ALIVE: '1800'
+      NCCL_NVLS_ENABLE: '1'
+      SGL_ENABLE_JIT_DEEPGEMM: 'false'
+      SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+      PYTHONNOUSERSITE: '1'
+      MC_INTRANODE_NVLINK: '1'
+      SGLANG_MOONCAKE_CUSTOM_MEM_POOL: INTRA_NODE_NVLINK
+    args:
+      served-model-name: Qwen/Qwen3.5-397B-A17B-FP8
+      trust-remote-code: true
+      tensor-parallel-size: 4
+      expert-parallel-size: 1
+      data-parallel-size: 1
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+      mamba-ssm-dtype: bfloat16
+      attention-backend: trtllm_mha
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      page-size: 64
+      mem-fraction-static: 0.8
+      context-length: 262144
+      chunked-prefill-size: 16384
+      max-prefill-tokens: 16384
+      max-running-requests: 64
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      enable-metrics: true
+      enable-cache-report: true
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      disable-radix-cache: true
+      speculative-algorithm: NEXTN
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+      cuda-graph-max-bs-decode: 32
+      stream-interval: 50
+      enable-symm-mem: true
+      scheduler-recv-interval: 10
+      mamba-radix-cache-strategy: no_buffer
+      mamba-track-interval: 128
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+sbatch_directives:
+  mem: '0'
+srun_options:
+  mem: '0'
+  container-remap-root: ''
+telemetry:
+  enabled: true
+  collect_interval_ms: 1000
+  storage_subdir: power
+  required: true
+  startup_timeout_seconds: 120
+  request_timeout_seconds: 2
+  collector_join_timeout_seconds: 12
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    port: 9401
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  concurrencies:
+  - 32
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    SRT_MEASUREMENT_WINDOW_BENCHMARK_TYPE: custom
+    SRT_MEASUREMENT_WINDOW_CONCURRENCIES: '32'
+    SRT_MEASUREMENT_WINDOW_RESULT_ROOT: /logs
+    PORT: '8000'
+    IS_MULTINODE: 'true'
+    EVAL_FRAMEWORK: lm-eval
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126_256k
+    AIPERF_MAX_CONTEXT_LENGTH: '262144'
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: 'true'
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0'
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'sglang:'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c48-write-through-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c48-write-through-mtp.yaml
new file mode 100644
index 000000000..5197d3be5
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c48-write-through-mtp.yaml
@@ -0,0 +1,176 @@
+# Experimental; retain only measured improvements over the current Pareto frontier.
+schema: 2
+name: qwen35-b200-fp8-agentx-disagg-1p1d-p-tp4-d-tp4-hicache-c48-write-through-mtp
+model:
+  path: qwen3.5-fp8
+  container: lmsysorg/sglang:nightly-dev-cu13-20260918-20518d85
+  precision: fp8
+engine: sglang
+dynamo:
+  install: true
+  source:
+    rev: 805a77f053d807b0d8def5d27f674a6df0ed839e
+resources:
+  gpu_type: b200
+  gpus_per_node: 8
+slurm:
+  time_limit: '4:00:00'
+frontend:
+  type: dynamo
+  nginx_container: nginx-sqsh
+  nginx_session_affinity: true
+  nginx_session_affinity_header: X-Dynamo-Session-ID
+  enable_multiple_frontends: true
+  num_additional_frontends: 1
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: '1'
+    PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+  args:
+    router-mode: kv
+    router-session-affinity-ttl-secs: '3600'
+roles:
+  prefill:
+    nodes: 1
+    workers: 1
+    gpus: 4
+    env:
+      PYTHONUNBUFFERED: '1'
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+      SGLANG_ENABLE_SPEC_V2: '1'
+      TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+      SGLANG_TIMEOUT_KEEP_ALIVE: '1800'
+      NCCL_NVLS_ENABLE: '1'
+      SGL_ENABLE_JIT_DEEPGEMM: 'false'
+      SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+      PYTHONNOUSERSITE: '1'
+      MC_INTRANODE_NVLINK: '1'
+      SGLANG_MOONCAKE_CUSTOM_MEM_POOL: INTRA_NODE_NVLINK
+    args:
+      served-model-name: Qwen/Qwen3.5-397B-A17B-FP8
+      trust-remote-code: true
+      tensor-parallel-size: 4
+      expert-parallel-size: 1
+      data-parallel-size: 1
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+      mamba-ssm-dtype: bfloat16
+      attention-backend: trtllm_mha
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      page-size: 64
+      mem-fraction-static: 0.8
+      context-length: 262144
+      chunked-prefill-size: 16384
+      max-prefill-tokens: 16384
+      max-running-requests: 96
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      enable-metrics: true
+      enable-cache-report: true
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      disable-cuda-graph: true
+      enable-symm-mem: true
+      scheduler-recv-interval: 10
+      stream-interval: 50
+      mamba-radix-cache-strategy: extra_buffer
+      mamba-track-interval: 2048
+      enable-hierarchical-cache: true
+      # Combined KV/Mamba budget: 860 GB per TP4 prefill worker.
+      hicache-size: 215
+      hicache-io-backend: kernel
+      hicache-mem-layout: page_first
+      hicache-write-policy: write_through
+  decode:
+    nodes: colocate
+    workers: 1
+    gpus: 4
+    env:
+      PYTHONUNBUFFERED: '1'
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+      SGLANG_ENABLE_SPEC_V2: '1'
+      TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+      SGLANG_TIMEOUT_KEEP_ALIVE: '1800'
+      NCCL_NVLS_ENABLE: '1'
+      SGL_ENABLE_JIT_DEEPGEMM: 'false'
+      SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+      PYTHONNOUSERSITE: '1'
+      MC_INTRANODE_NVLINK: '1'
+      SGLANG_MOONCAKE_CUSTOM_MEM_POOL: INTRA_NODE_NVLINK
+    args:
+      served-model-name: Qwen/Qwen3.5-397B-A17B-FP8
+      trust-remote-code: true
+      tensor-parallel-size: 4
+      expert-parallel-size: 1
+      data-parallel-size: 1
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+      mamba-ssm-dtype: bfloat16
+      attention-backend: trtllm_mha
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      page-size: 64
+      mem-fraction-static: 0.88
+      context-length: 262144
+      chunked-prefill-size: 16384
+      max-prefill-tokens: 16384
+      max-running-requests: 96
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      enable-metrics: true
+      enable-cache-report: true
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      disable-radix-cache: true
+      speculative-algorithm: NEXTN
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+      cuda-graph-max-bs-decode: 48
+      stream-interval: 50
+      enable-symm-mem: true
+      scheduler-recv-interval: 10
+      mamba-radix-cache-strategy: no_buffer
+      mamba-track-interval: 128
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+sbatch_directives:
+  mem: '0'
+srun_options:
+  mem: '0'
+  container-remap-root: ''
+telemetry:
+  enabled: true
+  collect_interval_ms: 1000
+  storage_subdir: power
+  required: true
+  startup_timeout_seconds: 120
+  request_timeout_seconds: 2
+  collector_join_timeout_seconds: 12
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    port: 9401
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  concurrencies:
+  - 48
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    SRT_MEASUREMENT_WINDOW_BENCHMARK_TYPE: custom
+    SRT_MEASUREMENT_WINDOW_CONCURRENCIES: '48'
+    SRT_MEASUREMENT_WINDOW_RESULT_ROOT: /logs
+    PORT: '8000'
+    IS_MULTINODE: 'true'
+    EVAL_FRAMEWORK: lm-eval
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126_256k
+    AIPERF_MAX_CONTEXT_LENGTH: '262144'
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: 'true'
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0'
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'sglang:'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c64-write-through-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c64-write-through-mtp.yaml
new file mode 100644
index 000000000..ac6a00a1d
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c64-write-through-mtp.yaml
@@ -0,0 +1,176 @@
+# Short replay improves the published frontier; full official qualification remains required.
+schema: 2
+name: qwen35-b200-fp8-agentx-disagg-1p1d-p-tp4-d-tp4-hicache-c64-write-through-mtp
+model:
+  path: qwen3.5-fp8
+  container: lmsysorg/sglang:nightly-dev-cu13-20260918-20518d85
+  precision: fp8
+engine: sglang
+dynamo:
+  install: true
+  source:
+    rev: 805a77f053d807b0d8def5d27f674a6df0ed839e
+resources:
+  gpu_type: b200
+  gpus_per_node: 8
+slurm:
+  time_limit: '4:00:00'
+frontend:
+  type: dynamo
+  nginx_container: nginx-sqsh
+  nginx_session_affinity: true
+  nginx_session_affinity_header: X-Dynamo-Session-ID
+  enable_multiple_frontends: true
+  num_additional_frontends: 1
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: '1'
+    PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+  args:
+    router-mode: kv
+    router-session-affinity-ttl-secs: '3600'
+roles:
+  prefill:
+    nodes: 1
+    workers: 1
+    gpus: 4
+    env:
+      PYTHONUNBUFFERED: '1'
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+      SGLANG_ENABLE_SPEC_V2: '1'
+      TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+      SGLANG_TIMEOUT_KEEP_ALIVE: '1800'
+      NCCL_NVLS_ENABLE: '1'
+      SGL_ENABLE_JIT_DEEPGEMM: 'false'
+      SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+      PYTHONNOUSERSITE: '1'
+      MC_INTRANODE_NVLINK: '1'
+      SGLANG_MOONCAKE_CUSTOM_MEM_POOL: INTRA_NODE_NVLINK
+    args:
+      served-model-name: Qwen/Qwen3.5-397B-A17B-FP8
+      trust-remote-code: true
+      tensor-parallel-size: 4
+      expert-parallel-size: 1
+      data-parallel-size: 1
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+      mamba-ssm-dtype: bfloat16
+      attention-backend: trtllm_mha
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      page-size: 64
+      mem-fraction-static: 0.8
+      context-length: 262144
+      chunked-prefill-size: 16384
+      max-prefill-tokens: 16384
+      max-running-requests: 128
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      enable-metrics: true
+      enable-cache-report: true
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      disable-cuda-graph: true
+      enable-symm-mem: true
+      scheduler-recv-interval: 10
+      stream-interval: 50
+      mamba-radix-cache-strategy: extra_buffer
+      mamba-track-interval: 2048
+      enable-hierarchical-cache: true
+      # Combined KV/Mamba budget: 860 GB per TP4 prefill worker.
+      hicache-size: 215
+      hicache-io-backend: kernel
+      hicache-mem-layout: page_first
+      hicache-write-policy: write_through
+  decode:
+    nodes: colocate
+    workers: 1
+    gpus: 4
+    env:
+      PYTHONUNBUFFERED: '1'
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+      SGLANG_ENABLE_SPEC_V2: '1'
+      TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+      SGLANG_TIMEOUT_KEEP_ALIVE: '1800'
+      NCCL_NVLS_ENABLE: '1'
+      SGL_ENABLE_JIT_DEEPGEMM: 'false'
+      SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+      PYTHONNOUSERSITE: '1'
+      MC_INTRANODE_NVLINK: '1'
+      SGLANG_MOONCAKE_CUSTOM_MEM_POOL: INTRA_NODE_NVLINK
+    args:
+      served-model-name: Qwen/Qwen3.5-397B-A17B-FP8
+      trust-remote-code: true
+      tensor-parallel-size: 4
+      expert-parallel-size: 1
+      data-parallel-size: 1
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+      mamba-ssm-dtype: bfloat16
+      attention-backend: trtllm_mha
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      page-size: 64
+      mem-fraction-static: 0.88
+      context-length: 262144
+      chunked-prefill-size: 16384
+      max-prefill-tokens: 16384
+      max-running-requests: 128
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      enable-metrics: true
+      enable-cache-report: true
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      disable-radix-cache: true
+      speculative-algorithm: NEXTN
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+      cuda-graph-max-bs-decode: 64
+      stream-interval: 50
+      enable-symm-mem: true
+      scheduler-recv-interval: 10
+      mamba-radix-cache-strategy: no_buffer
+      mamba-track-interval: 128
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+sbatch_directives:
+  mem: '0'
+srun_options:
+  mem: '0'
+  container-remap-root: ''
+telemetry:
+  enabled: true
+  collect_interval_ms: 1000
+  storage_subdir: power
+  required: true
+  startup_timeout_seconds: 120
+  request_timeout_seconds: 2
+  collector_join_timeout_seconds: 12
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    port: 9401
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  concurrencies:
+  - 64
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    SRT_MEASUREMENT_WINDOW_BENCHMARK_TYPE: custom
+    SRT_MEASUREMENT_WINDOW_CONCURRENCIES: '64'
+    SRT_MEASUREMENT_WINDOW_RESULT_ROOT: /logs
+    PORT: '8000'
+    IS_MULTINODE: 'true'
+    EVAL_FRAMEWORK: lm-eval
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126_256k
+    AIPERF_MAX_CONTEXT_LENGTH: '262144'
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: 'true'
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0'
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'sglang:'
diff --git a/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c8-mtp.yaml b/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c8-mtp.yaml
new file mode 100644
index 000000000..2492edcae
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c8-mtp.yaml
@@ -0,0 +1,175 @@
+# Experimental; retain only measured improvements over the current Pareto frontier.
+schema: 2
+name: qwen35-b200-fp8-agentx-disagg-1p1d-p-tp4-d-tp4-hicache-c8-mtp
+model:
+  path: qwen3.5-fp8
+  container: lmsysorg/sglang:nightly-dev-cu13-20260918-20518d85
+  precision: fp8
+engine: sglang
+dynamo:
+  install: true
+  source:
+    rev: 805a77f053d807b0d8def5d27f674a6df0ed839e
+resources:
+  gpu_type: b200
+  gpus_per_node: 8
+slurm:
+  time_limit: '4:00:00'
+frontend:
+  type: dynamo
+  nginx_container: nginx-sqsh
+  nginx_session_affinity: true
+  nginx_session_affinity_header: X-Dynamo-Session-ID
+  enable_multiple_frontends: true
+  num_additional_frontends: 1
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: '1'
+    PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+  args:
+    router-mode: kv
+    router-session-affinity-ttl-secs: '3600'
+roles:
+  prefill:
+    nodes: 1
+    workers: 1
+    gpus: 4
+    env:
+      PYTHONUNBUFFERED: '1'
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+      SGLANG_ENABLE_SPEC_V2: '1'
+      TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+      SGLANG_TIMEOUT_KEEP_ALIVE: '1800'
+      NCCL_NVLS_ENABLE: '1'
+      SGL_ENABLE_JIT_DEEPGEMM: 'false'
+      SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+      PYTHONNOUSERSITE: '1'
+      MC_INTRANODE_NVLINK: '1'
+      SGLANG_MOONCAKE_CUSTOM_MEM_POOL: INTRA_NODE_NVLINK
+    args:
+      served-model-name: Qwen/Qwen3.5-397B-A17B-FP8
+      trust-remote-code: true
+      tensor-parallel-size: 4
+      expert-parallel-size: 1
+      data-parallel-size: 1
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+      mamba-ssm-dtype: bfloat16
+      attention-backend: trtllm_mha
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      page-size: 64
+      mem-fraction-static: 0.8
+      context-length: 262144
+      chunked-prefill-size: 16384
+      max-prefill-tokens: 16384
+      max-running-requests: 32
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      enable-metrics: true
+      enable-cache-report: true
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      disable-cuda-graph: true
+      enable-symm-mem: true
+      scheduler-recv-interval: 10
+      stream-interval: 50
+      mamba-radix-cache-strategy: extra_buffer
+      mamba-track-interval: 2048
+      enable-hierarchical-cache: true
+      hicache-size: 104
+      hicache-io-backend: kernel
+      hicache-mem-layout: page_first
+      hicache-write-policy: write_through_selective
+  decode:
+    nodes: colocate
+    workers: 1
+    gpus: 4
+    env:
+      PYTHONUNBUFFERED: '1'
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PIP_EXTRA_INDEX_URL: https://pypi.nvidia.com
+      SGLANG_ENABLE_SPEC_V2: '1'
+      TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '1800'
+      SGLANG_TIMEOUT_KEEP_ALIVE: '1800'
+      NCCL_NVLS_ENABLE: '1'
+      SGL_ENABLE_JIT_DEEPGEMM: 'false'
+      SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+      PYTHONNOUSERSITE: '1'
+      MC_INTRANODE_NVLINK: '1'
+      SGLANG_MOONCAKE_CUSTOM_MEM_POOL: INTRA_NODE_NVLINK
+    args:
+      served-model-name: Qwen/Qwen3.5-397B-A17B-FP8
+      trust-remote-code: true
+      tensor-parallel-size: 4
+      expert-parallel-size: 1
+      data-parallel-size: 1
+      quantization: fp8
+      kv-cache-dtype: fp8_e4m3
+      mamba-ssm-dtype: bfloat16
+      attention-backend: trtllm_mha
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      page-size: 64
+      mem-fraction-static: 0.8
+      context-length: 262144
+      chunked-prefill-size: 16384
+      max-prefill-tokens: 16384
+      max-running-requests: 32
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      enable-metrics: true
+      enable-cache-report: true
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      disable-radix-cache: true
+      speculative-algorithm: NEXTN
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+      cuda-graph-max-bs-decode: 8
+      stream-interval: 50
+      enable-symm-mem: true
+      scheduler-recv-interval: 10
+      mamba-radix-cache-strategy: no_buffer
+      mamba-track-interval: 128
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+sbatch_directives:
+  mem: '0'
+srun_options:
+  mem: '0'
+  container-remap-root: ''
+telemetry:
+  enabled: true
+  collect_interval_ms: 1000
+  storage_subdir: power
+  required: true
+  startup_timeout_seconds: 120
+  request_timeout_seconds: 2
+  collector_join_timeout_seconds: 12
+  dcgm_exporter:
+    container_image: dcgm-exporter
+    port: 9401
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  concurrencies:
+  - 8
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    SRT_MEASUREMENT_WINDOW_BENCHMARK_TYPE: custom
+    SRT_MEASUREMENT_WINDOW_CONCURRENCIES: '8'
+    SRT_MEASUREMENT_WINDOW_RESULT_ROOT: /logs
+    PORT: '8000'
+    IS_MULTINODE: 'true'
+    EVAL_FRAMEWORK: lm-eval
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126_256k
+    AIPERF_MAX_CONTEXT_LENGTH: '262144'
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: 'true'
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0'
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'sglang:'
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 36e171688..06e8c65fc 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8249,3 +8249,129 @@ dsv41flash-fp4-h200-sglang-agentic-dspark:
     - dram-utilization: 0.80
       search-space:
       - { tp: 8, ep: 8, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16] }
+
+qwen3.5-fp8-b200-dynamo-sglang-agentic-disagg-mtp:
+  image: lmsysorg/sglang:nightly-dev-cu13-20260918-20518d85
+  model: Qwen/Qwen3.5-397B-A17B-FP8
+  model-prefix: qwen3.5
+  runner: cluster:b200-nscale
+  precision: fp8
+  framework: dynamo-sglang
+  router:
+    name: dynamo-router
+    version: 805a77f053d807b0d8def5d27f674a6df0ed839e
+  kv-p2p-transfer: mooncake
+  multinode: true
+  disagg: true
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.8
+      search-space:
+      - spec-decoding: mtp
+        kv-offloading: dram
+        conc-list:
+        - 8
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - CONFIG_FILE=recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c8-mtp.yaml
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+        kv-offload-backend:
+          name: hicache
+      - spec-decoding: mtp
+        kv-offloading: dram
+        conc-list:
+        - 16
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - CONFIG_FILE=recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c16-mtp.yaml
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+        kv-offload-backend:
+          name: hicache
+      - spec-decoding: mtp
+        kv-offloading: dram
+        conc-list:
+        - 24
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - CONFIG_FILE=recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c24-mtp.yaml
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+        kv-offload-backend:
+          name: hicache
+      - spec-decoding: mtp
+        kv-offloading: dram
+        conc-list:
+        - 32
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - CONFIG_FILE=recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c32-mtp.yaml
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+        kv-offload-backend:
+          name: hicache
+      - spec-decoding: mtp
+        kv-offloading: dram
+        conc-list:
+        - 48
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - CONFIG_FILE=recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c48-write-through-mtp.yaml
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+        kv-offload-backend:
+          name: hicache
+      - spec-decoding: mtp
+        kv-offloading: dram
+        conc-list:
+        - 64
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - CONFIG_FILE=recipes/qwen3.5/sglang/b200-fp8/agentx/disagg-1p1d-p-tp4-d-tp4-hicache-c64-write-through-mtp.yaml
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+        kv-offload-backend:
+          name: hicache
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 6b5e011b1..258691d3d 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8423,3 +8423,15 @@
     - "Enable required measured power for the complete H200 Kimi-K3 balanced recipe through the shared recipe-gated launcher."
     - "为 H200 Kimi-K3 balanced 完整配方启用必需实测功耗。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3053
+
+- config-keys:
+  - qwen3.5-fp8-b200-dynamo-sglang-agentic-disagg-mtp
+  scenario-type:
+  - agentic-coding
+  description:
+  - Add B200 Qwen3.5-397B-A17B FP8 AgentX disaggregation with colocated TP4 prefill
+    and TP4 decode on eight GPUs at concurrency 8, 16, 24, 32, 48 and 64.
+  - Use Dynamo/SGLang nightly 20260918-20518d85 with NEXTN MTP and Mooncake NVLink transfers.
+  - Enable 104 GB per-rank prefill HiCache at C8-C32; use 215 GB with first-use
+    write-through and decode static memory 0.88 at C48/C64.
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3269
diff --git a/runners/launch_b200-nscale-slurm.sh b/runners/launch_b200-nscale-slurm.sh
index cc24c382a..0c97d696a 100755
--- a/runners/launch_b200-nscale-slurm.sh
+++ b/runners/launch_b200-nscale-slurm.sh
@@ -545,6 +545,7 @@ run_multinode_srt() {
     fi
 
     USES_DCGM_POWER=0
+    USES_AGENTX_POWER=0
     _POWER_CONFIG_FILE="${CONFIG_FILE:-}"
     if [[ "${EVAL_ONLY}" == "true" && -n "${EVAL_CONFIG_FILE:-}" ]]; then
         _POWER_CONFIG_FILE="$EVAL_CONFIG_FILE"
@@ -560,13 +561,16 @@ run_multinode_srt() {
     ' "$_RECIPE_SRC"; then
         USES_DCGM_POWER=1
     fi
-    if [[ "$USES_DCGM_POWER" == "1" && (
+    if [[ "$USES_DCGM_POWER" == "1" && "$IS_AGENTIC" == "1" &&
+        "$MODEL_PREFIX" == "qwen3.5" && "$PRECISION" == "fp8" && "$FRAMEWORK" == "dynamo-sglang" ]]; then
+        USES_AGENTX_POWER=1
+    elif [[ "$USES_DCGM_POWER" == "1" && (
         "${IS_AGENTIC}" == "1" ||
         "$MODEL_PREFIX" != "dsv4" ||
         "$PRECISION" != "fp4" ||
         "$FRAMEWORK" != "dynamo-vllm"
     ) ]]; then
-        echo "Error: B200 Nscale dcgm-power is limited to fixed-sequence DSV4 FP4 dynamo-vllm" >&2
+        echo "Error: B200 Nscale dcgm-power requires fixed-sequence DSV4 FP4 dynamo-vllm or Qwen3.5 FP8 AgentX dynamo-sglang" >&2
         exit 1
     fi
 
@@ -678,8 +682,10 @@ run_multinode_srt() {
     sed -i 's/^  max_attempts: [0-9]*/  max_attempts: 720/' "${CONFIG_FILE%%:*}"
 
     SRTCTL_PREFLIGHT_ARGS=()
-    # Kimi K2.6 weights are staged on the Slurm compute nodes, not the login node.
-    if [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "kimik2.6" && $PRECISION == "fp4" ]]; then
+    # These weights are staged on the Slurm compute nodes, not the login node.
+    # SRT still checks the resolved model path when the worker starts.
+    if [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "kimik2.6" && $PRECISION == "fp4" ]] ||
+       [[ $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "qwen3.5" && $PRECISION == "fp8" ]]; then
         SRTCTL_PREFLIGHT_ARGS+=(--no-preflight)
     fi
 
@@ -700,29 +706,11 @@ run_multinode_srt() {
     LOGS_DIR="outputs/$JOB_ID/logs"
     LOG_FILE="$LOGS_DIR/sweep_${JOB_ID}.log"
 
-    while ! ls "$LOG_FILE" &>/dev/null; do
-        if ! squeue -j "$JOB_ID" --noheader 2>/dev/null | grep -q "$JOB_ID"; then
-            echo "ERROR: Job $JOB_ID failed before creating log file"
-            scontrol show job "$JOB_ID"
-            exit 1
-        fi
-        echo "Waiting for JOB_ID $JOB_ID to begin and $LOG_FILE to appear..."
-        sleep 5
-    done
-
-    (
-        while squeue -j "$JOB_ID" --noheader 2>/dev/null | grep -q "$JOB_ID"; do
-            sleep 10
-        done
-    ) &
-    POLL_PID=$!
-
-    echo "Tailing LOG_FILE: $LOG_FILE"
-
-    # -F follows by name and polls; inotify does not work on NFS.
-    tail -F -s 2 -n+1 "$LOG_FILE" --pid=$POLL_PID 2>/dev/null
-
-    wait $POLL_PID
+    local srt_job_rc=0
+    stream_slurm_job_log "$JOB_ID" "$LOG_FILE" || srt_job_rc=$?
+    if [[ "$srt_job_rc" -eq 0 ]]; then
+        verify_slurm_job_status "$JOB_ID" || srt_job_rc=$?
+    fi
 
     set -x
 
@@ -736,6 +724,15 @@ run_multinode_srt() {
 
     echo "Found logs directory: $LOGS_DIR"
 
+    if [[ "$USES_AGENTX_POWER" == "1" && "${EVAL_ONLY}" != "true" ]]; then
+        check_env_vars CONC_LIST
+        local -a power_concurrencies
+        read -r -a power_concurrencies <<< "$CONC_LIST"
+        collect_agentic_power_results "$JOB_ID" "$LOGS_DIR" \
+            "$GITHUB_WORKSPACE" "$GITHUB_WORKSPACE" "$RESULT_FILENAME" \
+            "$SRT_SLURM_COMMIT" "${power_concurrencies[@]}" || srt_job_rc=$?
+    fi
+
     if [[ "$USES_DCGM_POWER" == "1" ]]; then
         mkdir -p "$LOGS_DIR/power"
         cp "$GITHUB_WORKSPACE/exporter-image.sha256" "$LOGS_DIR/power/exporter-image.sha256"
@@ -776,6 +773,8 @@ run_multinode_srt() {
         sleep 10
     done
     find . -name '.nfs*' -delete 2>/dev/null || true
+    # Failed runs still provide the diagnostics and eval outputs above.
+    return "$srt_job_rc"
 }
 
 # ---------------------------------------------------------------------------
diff --git a/runners/slurm_utils.sh b/runners/slurm_utils.sh
index edb6d2a66..67a6eac78 100644
--- a/runners/slurm_utils.sh
+++ b/runners/slurm_utils.sh
@@ -157,6 +157,32 @@ stream_slurm_job_log() {
     wait "$poll_pid"
 }
 
+verify_slurm_job_status() {
+    local job_id="$1"
+    # Disappearing from squeue means terminal, not successful. Accounting can
+    # lag briefly; inspect only the allocation, never successful service steps.
+    local attempt accounting state exit_code
+    for attempt in {1..10}; do
+        accounting=$(sacct -X -n -P -j "$job_id" --format=State,ExitCode 2>/dev/null) || accounting=""
+        IFS='|' read -r state exit_code <<< "$accounting"
+        case "$state" in
+            COMPLETED)
+                if [[ "$exit_code" == "0:0" ]]; then
+                    return 0
+                fi
+                ;;
+            ""|PENDING|RUNNING|CONFIGURING|COMPLETING)
+                sleep 1
+                continue
+                ;;
+        esac
+        echo "ERROR: Slurm job $job_id ended with state=$state exit_code=$exit_code" >&2
+        return 1
+    done
+    echo "ERROR: could not verify terminal Slurm status for job $job_id" >&2
+    return 1
+}
+
 copy_to_workspace() {
     local source_file="$1"
     local destination_file="$2"