Category 05

AI Observability & Evaluation

LLM monitoring, tracing, testing, evaluations and guardrails.

Standout picks across editions

The current edition's pick, highlighted, plus earlier picks in this category from our backfilled archive — sourced from credible outlets (Hacker News, GitHub, TechCrunch and more) and checked against each product's official site.

This edition's pick · Nov 16–30, 2026
L

Langfuse

Open-source LLM observability & evaluation

Open-source tracing and evals for LLM apps.

Why it matters

Langfuse keeps shipping — this edition's evaluation-engine changes (LLM-as-a-judge targeting observations/experiments) reinforce its lead as the default open-source LLMOps layer.

Oct 16–31, 2026 edition
S

Sphinx

Quizzes you on your AI's code before shipping

Git hook that quizzes you on AI-written code.

Why it mattered

A hook that quizzes you on your AI's code before you ship it — a human-in-the-loop guardrail against merging code you don't understand.

Evidence dated 2026-10-18hackernews ↗
Researched · pending approvalVisit site ↗
Oct 1–15, 2026 edition
AB

Agent Box

Sandbox that restricts AI agents

OS-isolated, default-deny sandbox for running AI agents.

Why it mattered

Runs AI agents in an OS-isolated, default-deny sandbox with credential injection that keeps secrets out of the agent — guardrails for autonomous execution.

Evidence dated 2026-10-02github ↗
Researched · pending approvalVisit site ↗
Aug 16–31, 2026 edition
R

Reverify

Grounded, hallucination-checked agent toolkit

Agents propose; deterministic tools decide, every claim checked.

Why it mattered

An MCP server + CLI that makes agents propose while deterministic tools decide, checking every claim against ground truth — a direct answer to hallucination in agent workflows.

Evidence dated 2026-08-31github ↗
Researched · pending approvalVisit site ↗

How picks are selected

Candidates are scored on market traction, recent momentum, product innovation, community interest, practical usefulness and source credibility. Only researched, fact-checked products are published — if no candidate meets the evidence threshold, we show no pick rather than forcing a weak selection.