Knowledge catalogue
Learn how agent traces, human oversight, production evaluations, and log retention support EU AI Act compliance with a practical credit-assistant example. The post EU AI Act compliance for AI agents:
Build a reproducible benchmark, compare tool use across agents, and test whether a description change reduces unnecessary work while preserving outcomes. The post How to test agent tool descriptions w
Evaluate open models with Together AI in Arize AX. Compare model performance, run LLM-as-a-Judge evaluations, and explore lower-cost alternatives. The post Together AI integration: Evaluate open model
Automated evals can score every agent session, but they do not automatically reflect what your team considers a good outcome. That takes human judgment. The post Alignment is a human-AI workflow: how
AI costs don’t always scale predictably with traffic. A retry loop, oversized context window, or unnecessarily expensive model can make a small number of traces disproportionately expensive. The post
On 15 September, TypeSafe released Jev, and within a few weeks about 10 companies had shipped something that works the same way. OpenAI, Liquid, Cloudflare, AWS and PostHog all launched one, and an op
Some agent failures are already in the trace. tracelint proves them on Phoenix OpenInference spans and fails CI, without an LLM judge. The post What agent traces can tell you without an LLM judge appe
See how Arize uses Harbor and Phoenix to benchmark AI agents, MCP servers, skills, and CLIs in reproducible sandboxes, then inspect traces and annotations. The post How we benchmark AI agents and tool
Anthropic's Claude hillclimb loop designs evals and improves agents one change at a time. Here's how to run that loop against production traces in Arize AX. The post Claude’s hillclimb loop for AI age
MCP, CLI, and skills are three ways to extend an agent with the same platform. The question is no longer which one wins. It is where the agent runs, and who is asking. The post Introducing Arize AX MC
We benchmarked prompt caching across DeepSeek, GLM, GPT, and Claude using Harbor evals and Phoenix traces, so we could compare cache reuse, estimated cost, and latency on the same multi-turn shopping
Agents built and deployed on MongoDB’s Atlas Agent Engine can export OpenTelemetry traces to Arize AX, where teams can inspect each run, evaluate both the outcome and execution path, and test fixes be
This year agents at OpenAI, Anthropic, and Google escaped environments meant to contain them. What decided severity was time-to-detect and time-to-kill, from 12 minutes to seven months. NVIDIA's Open
Opus 5.5 dropped em dashes from 12.9 per 1,000 words to two in 57,000 words, and cut the rest of the Claudisms in half. Better, but not fixed. The post Anthropic says it fixed Claude’s writing. I ran
Harness distillation can train general scaffolding into a model. The harness that remains is the part tied to your tools, data, users, and environment. The post Are agent harnesses dying? What harness
After 237 talks across eight stages, code review emerged as the bottleneck for AI-generated code. Here are the production eval, sandbox, context, and agent experience themes that kept showing up. The
I tested Jev and five LLM judges across ten Arize Phoenix evaluators to measure how often their answers changed alongside accuracy, cost, and latency. The post What Jev’s probabilities reveal that rep
Evaluate agent responses with Jev and Arize AX. Build a FastAPI remote evaluator, return labels and scores, and choose a threshold for your data. The post How to build a remote evaluator in Arize AX w
We benchmarked Jev against Claude Opus 5 and GPT-5.6 Terra on accuracy, cost, and latency. Learn how threshold tuning changes hallucination detection. The post Jev vs. LLM-as-a-Judge: Accuracy and cos
Compare Jev and GPT-5.4 nano for real-time LLM guardrails, with demo results on latency, cost, and checks on agent inputs, replies, and tool calls. The post Real-time LLM guardrails with Jev: comparin
Sessions are now a first-class unit of work in Arize AX. Annotate a whole conversation, queue it for review, and ask Alyx to filter for it — plus Agent-as-a-Judge on every plan, vision judges, faster
TypeSafe’s Jev classifies, scores, and routes without generating text — up to hundreds of times cheaper than an LLM judge. What that changes for evals, confidence routing, and application architecture
Evals measure failures you know how to name. Arize Signal continuously reviews production traces to find recurring trajectory failures you do not, then turns them into evidence for tests and fixes. Th
A cancellation-policy fix raised action safety and dropped average task completion from 0.89 to 0.72. This walkthrough shows how to regression-test agent changes with Agent Experiments in Arize AX. Th
Code mode gives an agent a sandbox instead of a longer tool list. Here is why it fixes the too-many-tools problem, what it costs you in sandboxing and debuggability, and when plain tool calling is sti
By routing planning, exploration, implementation, and review to different models, I reduced one recurring coding-agent workflow from roughly 100 to 15-$20 per run. The post How I cut coding agent cost
By redesigning planning, validation, and risk review around AI agents, Coinbase Wallet dramatically shortened the path from product idea to working software. The post How Coinbase Wallet built an agen
Every LLM call your application makes costs money, and agentic applications make a lot of LLM calls. Arize AX now ships a Cost Agent that reads traces, ranks spend, and opens a pull request with the f
Trace where your AI budget goes, identify the requests that do not justify their spend, and validate lower-cost alternatives with Arize AX. The post How to reduce LLM costs without sacrificing quality
We ran Signal on Alyx, the AI engineering agent built into Arize AX. It surfaced a duplicate task-state loop and a 43-call dataset retry that appeared as valid tool activity or an OK root span. The po
In this installment of Rise of the AI Engineer, Stuart Sy from OpenAI, explains why the bottleneck has moved off the model and onto context, evals, and observability. The post Why better models don’t
A skill is just another AI agent: a prompt plus a harness that runs it. That means you can trace it, eval it, and prove a change made it better instead of shipping on vibes. Here's how we quantify ski
After Grok Build was caught uploading entire Git repos, we read the privacy docs for Claude Code, Codex, Cursor, GitHub Copilot, and Grok Build to compare what code leaves your machine, what gets reta
Agents changed what failure looks like, and the evaluation layer has to change with them. Why agent-as-a-judge is moving from research paper to production eval stack. The post Where agent evals are go
A background comment about pizza exposed a failure that Uber’s offline evaluations had missed. The incident helped reveal what production AI agent evaluation actually requires: automatic tracing, livi
Guardrails constrain what an agent can do in code; evals judge whether it performed well. Learn how both layers—and the harness around them—make long-running AI agents reliable. The post AI agent guar
Learn how evaluation-driven development, agent harnesses, AI observability, guardrails, and cost-per-outcome metrics move AI agents from pilot to production. The post Evaluation-driven development: Ho
Public benchmarks can show how a model performs in general. Production reliability depends on the context and harness around it, which only your team can evaluate against its own data, workflows, and
Orchestrator-executor just became the smart default for production agents: an expensive model plans, cheap models execute, and cost per completed task decides the roster. The post How cheap models cha
Traditional APM can collect every span and still leave developers guessing about intent, causality, and drift. As agents multiply, the observability stack must learn to interpret the systems it watche
Learn how Arize Signal turns production traces into ranked issues, proposed fixes, regression datasets, and reviewable pull requests for AI agents. The post How to debug production AI agents with Sign
Signal, a managed agent built into Arize AX, continuously reviews production traces, surfaces ranked issues with evidence and proposed fixes, and — with Managed Agents — can carry investigations into
Learn how to build trustworthy AI agent evals using regression tests, capability evals, production traces, LLM judges, and reproducible environments. The post AI agent evaluation: Tips from Anthropic
A skill cut agent costs by 44% and latency by 56%, but also reduced answer completeness. Here’s how tracing, evals, and a long-running agent exposed and corrected the regression. The post How to impro
How Booking.com built a unified AI observability stack with Arize for agentic GenAI workflows and traditional ML — from telemetry collection and PII redaction to latency monitors and evaluations. The
Three recent studies show what actually makes an AI agent skill effective: human expertise, compact procedures, tight routing, harness-specific testing, and eval-gated changes—not longer Markdown. The
Arize and Fireworks benchmarked 10 AI models across 2,400 agent runs. Learn why cost per successful task beats token price for model evaluation and routing. The post Cost per successful task: Benchmar
No single metric proves an LLM judge is trustworthy. This field guide shows how to measure human–human agreement, compare it to LLM–human agreement, and diagnose errors with precision, recall, and F1.
At ChatGPT scale, user frustration arrives as support tickets, ratings, social posts, and corrections buried inside conversations. OpenAI built a feedback system that can find the pattern behind a com
As background agents take on more implementation work, Cursor is rebuilding the software development lifecycle around risk scores, developer-like environments, video evidence, and review systems that
Use Arize Skills with Kiro CLI to trace coding-agent changes, build datasets from failures, run experiments, and validate prompts before shipping. The post Kiro CLI observability: trace and evaluate a
At Observe 2026, Jason Lopatecki and Aparna Dhinakaran described the shift from human-operated agent development to systematic agent improvement—and what builders should change in their stacks first.
AI productivity is best measured by connecting AI usage to validated downstream outcomes. Tokens, prompts, and generated lines show activity, but they do not prove value. A better measurement model tr
Microsoft published a 109-page technical report on MAI-Thinking-1. Here’s the abbreviated version of how a modern lab actually trains a frontier reasoning model — from scraping the web to reinforcemen
A local coding agent, an in-app customer assistant, and an AI SRE triaging production logs may all use the same model class—but not the same harness, eval plan, or rollout risk. Mastra CEO Sam Bhagwat
The AI engineering world is using “loop” to describe several different agent architectures. This post maps execution loops, task loops, product loops, system loops, and the human oversight loop that c
Kubernetes is strong for long-lived services, but it is often a poor default for short-lived agent sandboxes. Trace sandbox creation, tool execution, eval latency, and full trajectory time before you
WorkOS founder Michael Grinich explains why the next era of AI engineering depends on the systems around agents: identity, permissions, evals, memory, and feedback loops that keep autonomous software
If you're struggling to get started with evals, you're not alone. This post explains how to write LLM evals as ordinary tests in CI with Phoenix, pytest, and Vitest/Jest. The post Evals in CI: How to
As models become cheaper and more interchangeable, the durable advantage shifts to the agent harness: the loop, tools, memory, permissions, and workflow you can own and refine. The post Own the loop: