AgentSight brings local observability to AI agent trajectories
InfoQ reports that AgentSight is the observability component of Agentic OS and runs on the user’s own machine, where it reconstructs agent trajectories and analyzes their cost, performance, and accuracy. Product documentation cited by InfoQ says the current version supports Claude Code, Codex CLI, and Qwen Code, with macOS and Linux local session collection and optional full eBPF tracing on Linux. In AgentSight’s own frontend case study, adding a reusable AGENTS.md rule reportedly reduced tokens from 36K to 8.9K, LLM steps from 18 to 8, and elapsed time from 370 to 130 seconds; this is a vendor case rather than an independent measurement.
AgentSight provides a local workflow for tracing agent behavior, identifying avoidable turns, and preserving reusable project rules, while its reported savings still require independent validation.
背景
Agent 的 Token 花销里，有不少是冤枉钱。同一个项目跑过的坑，换一次任务又踩一遍；上次查清楚的信息，这次从头再查一轮。这些轮次本来不该发生，但每一步都在烧 Token。
业界主流的降本法是剪单次。工具返回太长就压，Headroom（headroomlabs-ai/headroom）压冗余结构，rtk（rtk-ai/rtk）过滤 shell 输出；历史太长就截，Claude Code 的 compaction、LangChain 的 trim_messages 做滑动窗口和摘要。每次交互少发一点，确实省钱。但这些手段对付的是单次调用的体积，管不了轮次本身该不该发生。
AgentSight 做的是后面这一层。先看清钱花在哪，再从轨迹里把经验攒下来，让不该发生的轮次别发生。跟剪单次互补，不冲突。
#AgentSight#Agentic OS#Claude Code#Codex CLI#Qwen Code#OpenTelemetry#Kubernetes#Docker