Arize benchmarked eight decision models — including TypeSafe's Jev 1.13, Liquid AI's Liquid d1, OpenAI Decisions (gpt-6-luna), Cloudflare's Clef 27B, and Jared Palmer's open Kev 9B — on hallucination …
Anthropic's Lance Martin published "Automating eval design and hillclimbing with Claude" on September 28, adding two commands to the claude-api skill: /claude-api build-eval, which interviews the user…
A comparison of four AI agent observability and evaluation tools — Langfuse, Arize Phoenix, Promptfoo, and Iris — finds they solve the same problem in fundamentally different ways: Langfuse and Phoeni…
Arize shipped long-term memory for Alyx, its AI engineering agent inside Arize AX, as a single structured text file per user and space capped at 8,000 characters that the agent reads in full on every …
Arize AX has added long-term memory to Alyx, its AI engineering agent, enabling it to retain project goals, conventions, preferences and prior decisions across sessions, and the feature is enabled by …
Arize's 2026 cost model breaks production evaluation spend into traffic volume × sampling rate × evaluation surfaces × evaluator cost plus human review and retention, while offline evaluation is datas…
A Peking University research team led by Haoran Ye trained the behavior of a specialized agent harness into Qwen3.5-9B via a method called harness distillation, raising the small model's average task …
Anthropic's Claude Opus 5.5 cut em dash usage to 2 occurrences across 57,000 words of output, down from 12.9 per 1,000 words in Opus 5, according to an independent evaluation run by Arize employee Jim…
Arize AX now integrates directly with TypeSafe AI, bringing native Jev-as-a-Judge evaluations into the Arize evaluation workflow, the company announced. In an Arize benchmark for hallucination detecti…
Arize published a tutorial showing how to connect TypeSafe's Jev to Arize AX through a remote evaluator that scores recorded agent spans after the agent responds, returning a yes/no label and a probab…
A benchmark of 23,325 judgments across the RAGTruth and SummEval human-labeled datasets found that TypeSafe's Jev matched Claude Opus 5 at 87% accuracy on held-out hallucination detection at roughly 1…
TypeSafe AI's purpose-built decision model Jev returned guardrail decisions faster and cheaper than OpenAI's GPT-5.4 nano in a dealership chatbot demo, according to a comparison published by Arize. Th…
Daytona cofounder and CEO Ivan Burazin said the agents he uses fail more often because they cannot access the tools they need than because of the model itself, and that he does not want agents operati…
A developer built COGEXT, a verifier engine designed to catch AI agents that falsely claim actions succeeded, such as sending an email or deploying code. The system extracts commitments from agent out…
Anthropic's Claude Code and Codex now have non-engineering variants in Claude Design and Codex Work, and Anthropic's revenue has been growing 10x year over year, according to a blog post by Arize. The…
A developer's debugging framework for AI agent observability identifies memory state, tool execution, and retrieval validity as the three critical pillars for diagnosing failures in production. The fr…
Stuart Sy, a member of OpenAI's technical staff, says that roughly one in every 30 to 50 agent conversations fails silently, and he argues that model intelligence is not the bottleneck; instead, failu…
AgentCheck, a new open-source regression testing tool for AI agents, lets developers define expected agent behavior in YAML and run pass/fail checks against any CLI command or HTTP endpoint, with an L…
This week's COSS newsletter reports that Stripe is in talks to acquire OpenRouter for over $7 billion, Databricks acquired Electric, and Dynatrace acquired Arize. Temporal is reportedly raising $500 m…
A research paper proposing Agent-as-a-Judge, an evaluation method where an AI agent assesses another agent's behavior, was accepted to ICML 2025 and began shipping in products by July 2026, moving fro…