AgentX
CC Traces Weka 062126
Real Claude Code agentic conversation traces captured for the InferenceX agentic benchmark. Each conversation is a sequence of turns (and subagent groups) with per-turn input/output token counts and 64-token KV-block hash IDs used to compute prefix-cache reuse.
Conversation overview
- Conversations
- 393
- Median requests / convo
- 86
- Mean requests / convo
- 251.5
- Main turns
- 56.8k
Agent and token mix
- Median subagents / trace
- 0
- Mean subagents / trace
- 4.3
- Cached input
- 98%
- Total tokens
- 21.7B
Model mix (turns)
claude-opus-4-8 62.1kclaude-fable-5 16.2kclaude-haiku-4-5-20251001 8.0kclaude-opus-4-7 6.4kclaude-opus-4-6 5.5kclaude-sonnet-4-6 576claude-sonnet-4-5 107
Distributions
Input tokens per turnlog scale
n=98,827 · p50 142.0k · p75 310.5k · p90 549.5k · p95 682.9k · max 989.8k tokens
Output tokens per turnlog scale
n=98,827 · p50 444 · p75 1.1k · p90 2.7k · p95 4.3k · max 59.9k tokens
Uncached input tokens per requestlog scale
n=98,827 · p50 1.7k · p75 2.9k · p90 6.8k · p95 14.7k · max 949.4k tokens
Turns per conversation
n=393 · p50 65 · p75 140 · p90 330 · p95 553 · max 3.0k turns
Subagent request ISLlog scale
Inner subagent requests only
n=42,029 · p50 64.0k · p75 119.7k · p90 196.3k · p95 259.5k · max 918.5k tokens
Subagent request OSLlog scale
Inner subagent requests only
n=42,029 · p50 328 · p75 640 · p90 1.3k · p95 2.0k · max 59.9k tokens
Cached fraction per turn
n=98,827 · p50 99% · p75 100% · p90 100% · p95 100% · max 100%
Conversations (0)
| Conversation | Turns | Subagents | Input | Output | Cached |
|---|---|---|---|---|---|
| Loading conversations… | |||||