[Feature Request] Support AgentX Workload Trace Splitting and Analysis

[jiaqiang-dot-liu]

1. AgentX Workload Overview

AgentX is the InferenceX benchmark for long-context, multi-turn coding-agent inference. It uses AIPerf to replay privacy-preserving workload shapes derived from real coding-agent sessions.

Unlike synthetic serving workloads with fixed input and output lengths, AgentX preserves:

• multi-turn session structure;
• growing shared prefixes and KV-cache reuse;
• dynamic input and output lengths;
• pauses between turns;
• parallel subagent branches and dependencies;
• closed-loop concurrency, where each lane represents an active agent session rather than an independent request.

The AgentX v1.0 dataset contains 393 sessions and includes contexts of up to approximately one million tokens. Request content is replaced with deterministic synthetic tokens, while request lengths, timing, prefix relationships, and branch structure are preserved.

The workload is executed through the following AIPerf scenario:

--scenario inferencex-agentx-mvp

This issue concerns TraceLens support for profiler traces produced while serving AgentX traffic. It is separate from the TraceLens Agent's own agentic report-generation workflow.

References:

• AgentX definition
• AgentX methodology
• Agentic benchmark metrics and methodology
• AIPerf AgentX MVP tutorial
• AIPerf Weka trace replay

2. Required TraceLens Adaptations

TraceLens needs the following capabilities for AgentX workloads:

• AgentX workload support: Recognize how AgentX differs from fixed-length synthetic serving workloads, including multi-turn sessions, dynamic lengths, shared prefixes, and parallel subagents.
• Steady-state selection and trace splitting: Select representative stable execution regions from AgentX traces and correctly handle short and multi-rank traces.
• Trace quality validation: Detect incomplete, unrepresentative, or otherwise unreliable traces and return an explicit status.
• Result aggregation: Combine analysis across AgentX phases and dynamic shapes into workload-level hotspot findings.
• Unified interface: Provide a complete entry point for steady-state selection, trace splitting, and analysis, requiring callers to supply only the traces and AgentX workload context.