AI observability has emerged as one of the biggest challenges in deploying and managing AI technology. Here are 10 AI observability and governance tools, from startups and leading technology providers, that solution providers should know.
All Eyes On AI
Businesses and organizations are developing and deploying AI systems—large language models, AI applications and AI agents—at a breakneck pace. But are these technologies effectively and safely doing what they are built to do?
Are they correctly producing the desired outputs or are they making errors and generating hallucinations? Are LLMs incorrectly accessing and/or misusing data and so putting a company at risk? And are they operating in a cost-efficient manner or are they rapidly burning AI tokens—and the organization’s AI budget in the process?
Many companies developing AI systems and running them in production are flying blind when it comes to answering these questions and this lack of observability is hindering the adoption and rollout of AI systems. AI can be a black box across all stages of the AI software lifecycle including development, testing, evaluation, deployment, monitoring and governance.
All this is generating demand for new tools for observing, monitoring and governing AI technology including LLMs, applications and agents.
The observability tools available today for IT systems are relatively mature. But they are designed for such tasks as warning when an application’s performance is faltering, a server is about to fail or a network has too much traffic.
Tracking the performance of AI systems is a whole different story. Because wrong answers can damage user trust in AI systems, the observability focus is on understanding the AI’s reasoning process. That means tracking AI prompt performance, monitoring for hallucinations and illogical outputs, detecting data drift, and measuring other unique metrics.
AI systems also need to be monitored to be sure they are accessing the right data—not just for producing correct results, but to be sure they aren’t somehow improperly accessing sensitive data and violating data use compliance policies or government regulations. And with the more recent focus on AI token usage, there’s a growing need for monitoring and managing the cost of AI systems in production.
Solution providers looking for AI observability, monitoring and governance tools have a lot to choose from.
There are a number of startups that offer observability tools specifically for AI LLMs, agents and applications. Some of their offerings are broad-based platforms covering all stages of the AI software lifecycle while others provide tools focused on specific areas such as AI software development.
The mainstream observability IT companies, meanwhile, have also been active here, building AI observability, monitoring and governance capabilities into their platforms.
LogicMonitor, for example, has expanded the capabilities of its flagship Envision observability platform with agentic AIOps to provide unified visibility across the entire AI service stack, tracking and managing the health, performance and cost of production AI systems, applications and large language models.
The need for AI observability capabilities has also spurred acquisition activity. In January cloud database builder ClickHouse bought Langfuse, developer of an AI engineering platform that helps software teams track, test and improve LLM-based applications.
In April, Cisco Systems acquired startup Galileo Technologies and its Galileo AI observability, evaluation and production guardrail platform for ensuring accuracy, reliability and real-time protection in generative AI and agentic applications. Cisco plans to pair Galileo’s agentic AI observability and protection capabilities with its Splunk Observability platform.
Here’s a look at 10 AI observability and governance tools that have caught our eye in 2026 (so far).
Arize AX
Arize AX, developed by Arize AI of Berkeley, Calif., helps developers evaluate their large language model applications and autonomous AI agents as they build them and monitor those products for errors and issues once they are in production.
Arize AX is an enterprise-scale, fully managed SaaS platform that provides trace observability capabilities, based on OpenTelemetry and OpenInference standards, that track how multi-step agent workflows execute, how tools are called, and where latency or errors occur.
The system runs automated evaluations at scale to detect hallucination rates, toxicity and the overall correctness of AI responses. And real-time production monitoring functionality detects system failures, performance degradation and data drift.
A built-in prompt and experimentation IDE (integrated development environment) provides a workspace for designing, optimizing, testing and comparing new prompts or agent configurations against production datasets. Also available is the Alyx AI debugging assistant that automatically scans logs, debugs agent issues, and suggests code improvements.
In May, Deloitte Canada became an authorized Arize AI reseller, combining the Arize AX platform with its consulting expertise to help clients move their AI initiatives from research and development to full-scale production.
Arize also offers Arize Phoenix, a free source-available tool for local development and fast debugging.
Braintrust
Braintrust describes itself as the observability and evaluation layer for production-ready AI. The Braintrust platform assists engineering and product management teams as they monitor, evaluate, score and improve the performance of their large language model applications and AI agents.
Braintrust is used by developers to test, compare and continuously improve AI systems by systematically measuring output quality. The platform, which originally launched in December 2023, includes the Brainstore custom-built database that was purpose-built to quickly search and analyze heavy AI trace logging data.
In February, the San Francisco-based company launched Braintrust Topics, an AI-powered pattern discovery and active observability feature within the Braintrust platform that automatically groups and classifies production traces from AI agents and LLM applications according to recurring patterns.
Braintrust raised $80 million in February in a Series B funding round led by ICONIQ.
Comet Optik
Comet Optik, Comet’s large language model evaluation and observability platform, is designed to help developers test, evaluate and monitor AI applications and optimize their performance.
Comet describes Optik as a bridge between software engineering and data science, providing the ability to record, sort, search and understand each step an LLM application takes to generate a response.
Optik’s capabilities include LLM observability and tracing, automated LLM evaluation tests, and prompt and agent optimization—the latter for identifying performance bottlenecks and deploying AI with appropriate guardrails.
Comet, based in New York, debuted Optik in December 2024 and has continued to expand its capabilities. In June, Comet added cost intelligence features to the tool, providing engineers with complete visibility into Anthropic Claude Code and OpenAI ChatGPT Codex to track and optimize spending around those coding agents.
Covasant Agent Management Suite
Covasant Technologies positions its Covasant Agent Management Suite as a control plane for enterprise AI agents.
CAMS, launched in January, is an enterprise platform for end-to-end AI agent lifecycle management including building, orchestrating, governing, monitoring and scaling every agent within an organization “without losing control, compliance or context,” according to the company.
The platform’s Agent Registry creates an organization-wide catalog for every AI agent, whether it is CAMs native or externally built. The AI Agent Control Tower, meanwhile, provides real-time production observability for an entire agent estate and can audit every LLM call, enforce guardrails, and monitor associated costs.
And the platform’s Governance functionality manages access control, policy enforcement and audit logging to enforce compliance with such regulations as RBAC, ABAC, SSO and GDPR/HIPAA, both at build time and runtime.
In April Covasant, with dual headquarters in Plano, Texas and Hyderabad, India, partnered with Google Cloud in a collaboration intended to expand enterprise adoption of agentic AI systems based on Gemini Enterprise.
Datadog Agent Observability
Datadog is one of the long-time players in the observability market with its unified observability and security platform for cloud applications. The company has been expanding its reach into the AI space, initially with Datadog LLM Observability, now expanded into Datadog Agent Observability.
Datadog Agent Observability helps development teams evaluate, troubleshoot, improve and trace AI agents and large language model workflows. Core capabilities include trace-level visibility, automated quality evaluations, behavior clustering patterns, experiments and testing, and cost and performance monitoring.
On June 30, New York-based Datadog announced that it had acquired Adaptive ML, a frontier AI startup that is developing a “reinforcement learning operations” or RLOps platform that will enable enterprises to build, own, and deploy their own specialized agents and models. Adaptive ML will be combined with Datadog’s AI research operations, accelerating the company’s R&D efforts around AI models and agentic LLM post-training for observability.
Dynatrace AI Observability
Dynatrace is another established observability tech company that has been extending the capabilities of its core platform to monitor, analyze, optimize and secure generative AI applications and workflows, large language models and AI agents.
Dynatrace AI Observability’s functionality includes out-of-the-box analytics, auto instrumentation, targeted metrics and ready-made dashboards. Core capabilities include end-to-end distributed tracing, intelligent AI guardrails, cost and token governance, comprehensive stack monitoring, and compliance and auditing.
Boston-based Dynatrace initially launched AI Observability in 2024. Earlier this year the company significantly updated the product with an emphasis on autonomous operations and agentic ecosystems.
Dynatrace AI Observability supports more than 20 leading AI platforms including OpenAI, Amazon Bedrock, Google Gemini, Google Vertex, Anthropic and LangChain.
JetStream Security-First AI Governance Platform
JetStream Security describes its “security-first” AI governance (SAIG) platform as a control plane for AI.
JetStream provides enterprises with visibility into how AI systems behave in production through its real-time governance, identity, design control, and financial accountability capabilities.
The platform is anchored by JetStream Blueprints, what the company calls “living operational contracts,” that make every AI agent, model, workflow and identity within an enterprise visible, attributable and governed in real time. It continuously discovers AI systems across the organization, replaces raw credentials with virtual revocable keys, enforces runtime policy on every agentic action, and ties workflows to responsible owners.
JetStream, a startup headquartered in Santa Clara, Calif., was founded in 2025 by several former CrowdStrike and SentinelOne executives. In March of this year the company exited stealth with $34 million in seed funding from Redpoint Ventures and a number of angel investors including CrowdStrike CEO George Kurtz, Wiz CEO Assaf Rappaport and Okta co-founder Frederic Kerrest.
LangChain LangSmith Platform
LangSmith is LangChain’s commercial platform for prototyping, debugging, testing, evaluating, deploying and monitoring large language models, AI agents and AI workflows built using the open-source LangChain framework.
LangChain, founded in 2023, develops the LangChain framework (available under the MIT open-source license) for building AI applications and agents using LLMs, while also offering commercial products such as LangSmith.
LangSmith’s core functionality includes observability and tracing (trace-level debugging and multi-turn threads); evaluation and testing (dataset management and hybrid scoring); production monitoring (cost and latency tracking, error and feedback loops); managed deployment (durable runtime and agent studio); and advanced automation and agent runtime tools (including the LangSmith Engine and LangSmith Sandboxes).
LangSmith is framework-agnostic, meaning that it can be used outside of the LangChain universe and integrates with Python/TypeScript setups, the OpenAI SDK, Anthropic’s SDK or Vercel SDK.
In 2025 San Francisco-based LangChain raised $125 million in a Series B funding round that boosted the company’s valuation to $1.25 billion.
New Relic AI Observability
New Relic AI Observability provides end-to-end performance monitoring, cost governance and real-time troubleshooting for AI applications and agentic workflows. It is a core component of the company’s flagship New Relic Intelligent Observability Platform.
New Relic AI Observability’s capabilities include AI stack monitoring and integrations (full-stack tracing, ecosystem connectivity, and model performance mapping); agentic AI governance (multi-agent tracing, no-code agent builder, and Model Context Protocol); cost and risk management (token tracking, hallucination and drift detection, and enterprise guardrails); and generative troubleshooting and AIOps (natural language querying, root-cause reasoning, and log and anomaly analysis).
In June the San Francisco-based company introduced a new feature called New Relic AI Coding Observability that’s specifically designed for AI-assisted software development. It extends production-grade monitoring directly into the coding phase of the software lifecycle, providing visibility into a development organization’s use of AI coding assistants. The tool helps developers measure productivity gains from AI coding assistants and exercises cost controls over their use.
Pendo Agent Analytics
Pendo Agent Analytics provides a way to measure the performance, usage and business impact of AI agents across an enterprise. Built on the Pendo Software Experience Platform, Pendo Agent Analytics was released to general availability in December 2025 following a six-month early access trial period.
During that early trial, according to Pendo, dozens of organizations used the software to understand how users interact with AI agents, measure the quality of agent outputs, and identify where agent behavior supported or hindered user productivity.
Core functionality includes tracking hybrid workflows across both agents and traditional software, surfacing patterns in how people interact with agents, flagging “off-script” behavior such as hallucinations and inability to respond, and mapping agent activity to task completion and overall engagement for ROI and performance assessments.
In February, Raleigh, N.C.-based Pendo acquired Chisel Labs, a startup that develops an AI-powered platform that development teams use to write product requirement documents and release notes, conduct user research, and triage user feedback.
