AgentPProf：用 GPT-6 Sol/Luna 剖析智能体开销
阿里与 Eunomia Labs 提出语义剖析器 AgentPProf，在 GPT-6 Sol、Luna 与 Opus 5.5 上跑出 0.764 F1，问题定位 MAP 提升 56%。
this is pure f*cking fantastic Claude Opus 5.5 and GPT-6 Sol exposed the hidden agent tax Long-horizon agents break for a simple reason: unprofiled token bloat. In production runs lasting hours or days, naive prompt loops dump raw transcripts into massive context windows, causing prompt degradation and burning budgets on silent verification loops Researchers from Alibaba Group and Eunomia Labs introduced AgentPProf, a semantic profiler that formalizes agent activity as an Operation Stack replacing static call stacks with query-time intent projections. Running this system across GPT-6 Sol, GPT-6 Luna, and Claude Opus 5.5 sets a 0.764 B³ F1 score and raises problem-localization MAP by 56% Trajectory capture → GPT-6 Luna recursive segmentation → Operation stack projection → Cross-session pprof folding → Claude Opus 5.5 synthesis Four operational disciplines driving the semantic profiling engine: 1. Semantic Operation Stacks Agent activities across prompts, LLM calls, and kernel-level eBPF probes fold into uniform records. Query-time stack projections group scattered actions into standardized task paths across independent runs 2. Recursive Operation Segmentation GPT-6 Luna acts as a high-speed segmentation worker at $0.10 per million tokens. The model recursively cuts trajectories at task transition points to recover hierarchical intent without expensive turn-by-turn re-annotation 3. Differential Flame Graph Isolation Subtracting successful runs from failing ones isolates the root cause of crashes. Failing runs waste 44.6% of step budgets in stubborn recovery loops, retrying SSH variants instead of resolving broken credentials 4. Unified Context RAM and Memory SSD Byte-0 immutable system prompts maintain 99.47% KV-cache read rates on Claude Opus 5.5. Meanwhile, background memory loops apply Ebbinghaus exponential decay to prune stale entities before context contamination happens Unifying semantic profiling, deterministic harness boundaries, and dual-loop memory transforms fragile agent scripts into self-healing software infrastructure Read the complete breakdown in the article below ↓
这条是引用着另一条写的：x.com/marfinxx/status/2088234998654472340