Agentic coding workload
Also known as coding-agent workload, software-engineering agent workload
In plain English
This is the request pattern created when a coding agent reads a repository, edits code, runs tools, and revisits the model until the task is done.
Technical definition
An agentic coding workload is a multi-turn inference workload produced by a software agent that combines model generation with repository inspection, tool calls, code changes, and delegated subtasks.
Engineering details
Request sizes grow as the agent accumulates instructions, files, tool results, and earlier responses. Many turns reuse a large common prefix. Tool execution inserts uneven delays, while subagents can create overlapping request branches. These properties produce a different traffic shape from fixed-length prompt benchmarks.
Why it matters
Coding agents can keep a serving system busy for minutes or hours through a chain of dependent calls. Cache policy, scheduler fairness, memory capacity, and tail latency all affect task progress. Peak decode throughput alone cannot describe that behavior.
How to read it in InferenceX
AgentX represents this workload with trace-derived request shapes and deterministic synthetic content. It measures inference-system performance. Model coding quality requires a separate evaluation, so quality scores and AgentX serving results answer separate questions.
Source material
See the concept in real benchmarks
Ultra-High Interactivity on NVIDIA GPUs? TileRT on InferenceX
Can TileRT software on NVIDIA GPUs compete with Cerebras, Groq LPU, and SambaNova? Batch size 1, disaggregated engine, high-throughput prefill engine, high-interactivity decode engine
Vera Rubin NVL72 vs GB200 NVL72? Inference TCO & Architecture Analysis
Rubin LUT Based Tensor Core, Feynman, Rack Scale, Perf Per MegaWatt, Perf Per Dollar, Software Improvements, Public Rubin Software, PyTorch, vLLM, OpenAI Triton
InferenceX v2: NVIDIA Blackwell Vs AMD vs Hopper - Formerly InferenceMAX
GB300 NVL72, MI355X, B200, H100, Disaggregated Serving, Wide Expert Parallelism, Large Mixture of Experts, SGLang, vLLM, TRTLLM