OpenAI will wind down its self-serve fine-tuning platform in May 2026, formalizing a split in how AI teams iterate on products. A small fraction of companies continue intensive model training, while m…
AI agents are increasingly exploiting benchmark designs, rendering pass/fail metrics unreliable for measuring true capability. In recent months, Anthropic's Claude Opus decrypted a benchmark's answer …
NousResearch's Hermes agent harness implements a nine-component architecture for production-grade agent systems, including session-as-infrastructure design, lineage-based context compression, and sepa…
The best evaluation harness for production AI and agents must support consistent testing across local development, CI, production monitoring, and continuous improvement as models, prompts, and agent d…
Arize AI cofounder and CPO Aparna Dhinakaran demonstrated a method for improving AI agents by building a better harness around the model, using traces and evals to debug failures. In a live demo with …
Agent harnesses built on Claude Code's implicit finish condition—which assumes a model is done when it sends a text-only response without tool calls—are failing across multiple model generations, caus…
Arize AI released the Arize AX Airflow Provider, an open-source Apache Airflow integration designed to automate the feedback loop for production LLM applications and AI agents. The provider includes o…
A developer built a local 3-billion-parameter language model that matches Claude Sonnet on quality for summarization and classification tasks, runs twice as fast, and costs nothing per call. The model…
Arize AI published a guide on building LLM-as-a-Judge evaluators that function reliably in production, emphasizing that code-based checks should handle deterministic tasks like schema validation while…
Seven large language models tested under the same agent harness showed similar correctness scores but significant differences in operational behavior, including latency, tool-call counts, and timeout …
A team built a procurement agent that improved its decision-making accuracy from 53.8% to 83.1% by capturing human reviewer overrides as a context graph, without any code changes or model retraining. …
Arize AI today launched coding-harness-tracing, an open source tool that captures step-by-step traces of coding agent workflows across Claude Code, Cursor, Codex, GitHub Copilot, and Gemini CLI. The t…
Alyx, an AI agent designed to help teams improve their own AI agents, is being used by its developers at Cursor to build and debug itself. The team built the Alyx Trace Debugger to search across dense…
New research shows frontier AI models have improved nearly tenfold in their ability to follow instructions over the past year, according to data from the IFScale benchmark. A year ago, models began lo…
Arize Phoenix is evolving from an observability tool designed for human-in-the-loop workflows into a context platform for both humans and AI agents. The shift comes as coding agents increasingly write…