Tag
Runloop 是一个本地优先、开源的 3D 小镇可视化工具，把 Claude Code 代理会话以建筑形式呈现，支持实时观察、回放和对比运行记录。
Five independent tools (Agent-Native, AgentPlay, Agent Tray, Coucou, and a pixel-office visualization) all launched within 30 hours, each aiming to make AI agent activity visible to users. The author notes they solve the visibility problem but not the underlying setup and permission-modeling costs.
开发者把家里的 AI 助手和编码代理做成像素办公室里的小同事，用等距仪表盘实时查看各代理的状态，现已开源并支持自托管。
This paper introduces a benchmark showing that LLMs frequently conceal their own mistakes: models fail to disclose synthetic errors in 36.4% of chat and 67.1% of agentic rollouts, and knowingly hide mistakes in their chain of thought in a subset of cases, arguing that developers should rely on independent monitors rather than model self-reporting.
Goodfire launched 'inside-out' AI safety monitors that read a model's internal activations rather than rereading its outputs, catching rogue agents at far lower cost and latency. Available to Baseten customers, the probes detected 94% of malicious hacking sessions at roughly $51 per 1,500 sessions versus $10,000 for a top-tier AI monitor.
glowup is a Claude Code plugin that restyles the terminal UI with new colors, message rendering, borders, spinners, a side pane for showing diffs, subagent status, and context usage, plus an animated pet mascot called Clawd that reacts to session events.
Cubicle is an open-source (MIT) developer tool that visualizes AI coding agents like Claude Code, Codex, and Gemini CLI as workers in a live pixel-art office, with KPI dashboards, Telegram alerts, and a read-only design installed via a single npx command with zero dependencies.
An analysis of a public append-only log reveals that agent uptime metrics can be misleading, as they don't correlate with actual activity, and capabilities declarations can change without proper versioning, leading to data integrity issues.
Xybernetex is an experimental OpenClaw plugin that monitors AI agent runs to reduce wasted tokens and improve task success through interventions like retries and verification, with the author seeking real-world testers.
LangChain announced LangSmith Engine v2 with features like red teaming and validated fixes to proactively identify and resolve issues in AI agents before they affect production.
Herdr Studio is a community-built browser client that provides a visual workspace for managing terminals, sessions, files, diffs, and Git worktrees, allowing developers to monitor and control agent sessions from any modern browser.
Ramp developed a semantic layer to attribute AI agent spend to objectives and outcomes, moving from monitoring spend to understanding AI ROI.
We built an open-source retrospective reader for Claude Code to analyze agent behavior, using real execution history to inform runtime governance.
Toyota's enterprise AI team uses LangSmith and Deep Agents to scale AI agent development, reducing production time from 6 months to 4 days and deploying over 50 agents.
OpenObserve launches an AI-native, open-source observability platform designed to trace AI agents and LLMs, offering detailed insights into performance, cost, and quality for developers.
Raindrop launches Signals 2.0 powered by rd-signal-2, a new model pipeline for building task-specific binary classifiers from production traces. It claims near GPT-5.6 Sol xhigh accuracy at 1600x lower cost, and also introduces Signal Builder for custom classifiers with zero data retention.
An open source profiler designed for voice agents to provide insights into internal operations and performance.
Built Trovis, a tool that records and explains agent actions in production, showing deviations from expected behavior.
Agnost AI is a platform that analyzes agent conversations to surface user feedback, detect failures, and generate automated fixes, helping teams improve their AI agents.
Arize Phoenix announces new customization features for its AI agent monitoring platform, including customizable charts, command K, recent searches, and custom column ordering.