Arize AI launched Signal, a managed agent built into its Arize AX platform that continuously reviews production traces, identifies recurring failure patterns, groups them into ranked issues with root-…
An agent skill reduced average latency by 56%, token usage by 27%, and estimated cost by 44%, but also decreased answer completeness. The regression was detected only after tracing both versions in Ar…
Anthropic technical staff member Marius Buleandra reported that a newer AI model appeared to beat its predecessor by nine points on an AI data analyst eval, but the improvement vanished after he fixed…
Anthropic technical staff member Marius Buleandra warns that AI agent evaluations can produce misleading results, citing a case where a newer model appeared to beat its predecessor by nine points on a…
Booking.com relies on Arize AI for observability across its ecosystem of traditional ML models and generative AI agents, including trip planning assistants, ranking systems, and fraud detection models…
A good AI agent skill is a compact package of procedural expertise that fixes a repeatable failure, loads only when relevant, and earns its place against a matched evaluation, according to three recen…
Arize and Fireworks tested 10 AI models on 40 real agent tasks across 2,400 runs, finding that the open-source model gpt-oss-120b had the lowest cost per successful task at $0.054 despite a 33% pass r…
A new guide breaks down evaluation alignment between human experts and LLM judges into three measurable questions: human reliability on the rubric, LLM-human agreement relative to human-human agreemen…
OpenAI has built a feedback system that aggregates explicit and implicit user signals into a shared data layer, enabling teams to measure failures across channels and close an AI improvement loop. Stu…
Cursor treats verification as an integrated architecture across CI, security review, risk scoring, behavioral artifacts, and specialized review agents, with roughly 30 to 40 percent of pull requests i…
Amazon's Kiro CLI coding agent and Arize Skills now integrate to let developers trace, evaluate, and validate agent-generated code changes before merging, closing the gap between faster code generatio…
Arize cofounders Jason Lopatecki and Aparna Dhinakaran at Arize Observe 2026 outlined a shift from human-operated agent development to systematic, automated agent improvement, arguing that as agents s…
Arize AI, co-authored by Duncan McKinnon and Jitendra Yadav, argues that AI productivity should be measured by connecting AI usage to validated downstream outcomes rather than activity metrics like to…
Microsoft published a 109-page technical report on its MAI-Thinking-1 reasoning model, detailing the entire training process from web scraping to final optimization. The report reveals that 54.6% of t…
Sam Bhagwat, CEO of Mastra, identified three production patterns for AI agents—customer-facing, internal enterprise, and developer platform—each with distinct harness, eval, and rollout strategies. Sp…
The AI engineering community has adopted the term 'loop' to describe at least four distinct architectures: execution loops for agent decision-making, task loops that restart agents until tests pass, p…
WorkOS founder Michael Grinich warns that AI agents can fail in ways that look like success, such as deleting tests to make them pass. As agents become users, software must shift from human-centric de…
Agent sandbox requirements have outpaced Kubernetes defaults, making pod templates a bottleneck for short-lived agent workloads. Arize and Daytona CEO Ivan Burazin argue that teams should trace sandbo…
Arize Phoenix released evaluation support for Pytest and Vitest/Jest, allowing developers to write LLM evaluations as ordinary tests. The integration records LLM traces and metrics like cost and laten…
A new field guide maps agent harnesses—the software loops that let AI models edit files, run tests, and iterate autonomously—on capability and freedom. The analysis shows that as models become commodi…