AI observability has emerged as one of many greatest challenges in deploying and managing AI expertise. Listed here are 10 AI observability and governance instruments, from startups and main expertise suppliers, that answer suppliers ought to know.
All Eyes On AI
Companies and organizations are creating and deploying AI programs—giant language fashions, AI functions and AI brokers—at a breakneck tempo. However are these applied sciences successfully and safely doing what they’re constructed to do?
Are they appropriately producing the specified outputs or are they making errors and producing hallucinations? Are LLMs incorrectly accessing and/or misusing knowledge and so placing an organization in danger? And are they working in a cost-efficient method or are they quickly burning AI tokens—and the group’s AI price range within the course of?
Many corporations creating AI programs and operating them in manufacturing are flying blind in relation to answering these questions and this lack of observability is hindering the adoption and rollout of AI programs. AI could be a black field throughout all phases of the AI software program lifecycle together with growth, testing, analysis, deployment, monitoring and governance.
All that is producing demand for brand new instruments for observing, monitoring and governing AI expertise together with LLMs, functions and brokers.
The observability instruments obtainable in the present day for IT programs are comparatively mature. However they’re designed for such duties as warning when an software’s efficiency is faltering, a server is about to fail or a community has an excessive amount of visitors.
Monitoring the efficiency of AI programs is a complete totally different story. As a result of unsuitable solutions can harm person belief in AI programs, the observability focus is on understanding the AI’s reasoning course of. Which means monitoring AI immediate efficiency, monitoring for hallucinations and illogical outputs, detecting knowledge drift, and measuring different distinctive metrics.
AI programs additionally must be monitored to make sure they’re accessing the appropriate knowledge—not only for producing appropriate outcomes, however to make sure they aren’t someway improperly accessing delicate knowledge and violating knowledge use compliance insurance policies or authorities rules. And with the newer give attention to AI token utilization, there’s a rising want for monitoring and managing the price of AI programs in manufacturing.
Answer suppliers searching for AI observability, monitoring and governance instruments have so much to select from.
There are a selection of startups that supply observability instruments particularly for AI LLMs, brokers and functions. A few of their choices are broad-based platforms overlaying all phases of the AI software program lifecycle whereas others present instruments targeted on particular areas reminiscent of AI software program growth.
The mainstream observability IT corporations, in the meantime, have additionally been energetic right here, constructing AI observability, monitoring and governance capabilities into their platforms.
LogicMonitor, for instance, has expanded the capabilities of its flagship Envision observability platform with agentic AIOps to supply unified visibility throughout the complete AI service stack, monitoring and managing the well being, efficiency and price of manufacturing AI programs, functions and huge language fashions.
The necessity for AI observability capabilities has additionally spurred acquisition exercise. In January cloud database builder ClickHouse purchased Langfuse, developer of an AI engineering platform that helps software program groups observe, take a look at and enhance LLM-based functions.
In April, Cisco Techniques acquired startup Galileo Applied sciences and its Galileo AI observability, analysis and manufacturing guardrail platform for making certain accuracy, reliability and real-time safety in generative AI and agentic functions. Cisco plans to pair Galileo’s agentic AI observability and safety capabilities with its Splunk Observability platform.
Right here’s a take a look at 10 AI observability and governance instruments which have caught our eye in 2026 (up to now).
Arize AX
Arize AX, developed by Arize AI of Berkeley, Calif., helps builders consider their giant language mannequin functions and autonomous AI brokers as they construct them and monitor these merchandise for errors and points as soon as they’re in manufacturing.
Arize AX is an enterprise-scale, totally managed SaaS platform that gives hint observability capabilities, based mostly on OpenTelemetry and OpenInference requirements, that observe how multi-step agent workflows execute, how instruments are known as, and the place latency or errors happen.
The system runs automated evaluations at scale to detect hallucination charges, toxicity and the general correctness of AI responses. And real-time manufacturing monitoring performance detects system failures, efficiency degradation and knowledge drift.
A built-in immediate and experimentation IDE (built-in growth atmosphere) supplies a workspace for designing, optimizing, testing and evaluating new prompts or agent configurations towards manufacturing datasets. Additionally obtainable is the Alyx AI debugging assistant that mechanically scans logs, debugs agent points, and suggests code enhancements.
In Could, Deloitte Canada grew to become a certified Arize AI reseller, combining the Arize AX platform with its consulting experience to assist shoppers transfer their AI initiatives from analysis and growth to full-scale manufacturing.
Arize additionally provides Arize Phoenix, a free source-available instrument for native growth and quick debugging.
Braintrust
Braintrust describes itself because the observability and analysis layer for production-ready AI. The Braintrust platform assists engineering and product administration groups as they monitor, consider, rating and enhance the efficiency of their giant language mannequin functions and AI brokers.
Braintrust is utilized by builders to check, evaluate and constantly enhance AI programs by systematically measuring output high quality. The platform, which initially launched in December 2023, contains the Brainstore custom-built database that was purpose-built to rapidly search and analyze heavy AI hint logging knowledge.
In February, the San Francisco-based firm launched Braintrust Subjects, an AI-powered sample discovery and energetic observability characteristic inside the Braintrust platform that mechanically teams and classifies manufacturing traces from AI brokers and LLM functions in keeping with recurring patterns.
Braintrust raised $80 million in February in a Sequence B funding spherical led by ICONIQ.
Comet Optik
Comet Optik, Comet’s giant language mannequin analysis and observability platform, is designed to assist builders take a look at, consider and monitor AI functions and optimize their efficiency.
Comet describes Optik as a bridge between software program engineering and knowledge science, offering the flexibility to document, type, search and perceive every step an LLM software takes to generate a response.
Optik’s capabilities embody LLM observability and tracing, automated LLM analysis assessments, and immediate and agent optimization—the latter for figuring out efficiency bottlenecks and deploying AI with acceptable guardrails.
Comet, based mostly in New York, debuted Optik in December 2024 and has continued to broaden its capabilities. In June, Comet added value intelligence options to the instrument, offering engineers with full visibility into Anthropic Claude Code and OpenAI ChatGPT Codex to trace and optimize spending round these coding brokers.
Covasant Agent Administration Suite
Covasant Applied sciences positions its Covasant Agent Administration Suite as a management airplane for enterprise AI brokers.
CAMS, launched in January, is an enterprise platform for end-to-end AI agent lifecycle administration together with constructing, orchestrating, governing, monitoring and scaling each agent inside a corporation “with out shedding management, compliance or context,” in keeping with the corporate.
The platform’s Agent Registry creates an organization-wide catalog for each AI agent, whether or not it’s CAMs native or externally constructed. The AI Agent Management Tower, in the meantime, supplies real-time manufacturing observability for a whole agent property and might audit each LLM name, implement guardrails, and monitor related prices.
And the platform’s Governance performance manages entry management, coverage enforcement and audit logging to implement compliance with such rules as RBAC, ABAC, SSO and GDPR/HIPAA, each at construct time and runtime.
In April Covasant, with twin headquarters in Plano, Texas and Hyderabad, India, partnered with Google Cloud in a collaboration supposed to broaden enterprise adoption of agentic AI programs based mostly on Gemini Enterprise.
Datadog Agent Observability
Datadog is likely one of the long-time gamers within the observability market with its unified observability and safety platform for cloud functions. The corporate has been increasing its attain into the AI area, initially with Datadog LLM Observability, now expanded into Datadog Agent Observability.
Datadog Agent Observability helps growth groups consider, troubleshoot, enhance and hint AI brokers and huge language mannequin workflows. Core capabilities embody trace-level visibility, automated high quality evaluations, conduct clustering patterns, experiments and testing, and price and efficiency monitoring.
On June 30, New York-based Datadog introduced that it had acquired Adaptive ML, a frontier AI startup that’s creating a “reinforcement studying operations” or RLOps platform that may allow enterprises to construct, personal, and deploy their very own specialised brokers and fashions. Adaptive ML will probably be mixed with Datadog’s AI analysis operations, accelerating the corporate’s R&D efforts round AI fashions and agentic LLM post-training for observability.
Dynatrace AI Observability
Dynatrace is one other established observability tech firm that has been extending the capabilities of its core platform to observe, analyze, optimize and safe generative AI functions and workflows, giant language fashions and AI brokers.
Dynatrace AI Observability’s performance contains out-of-the-box analytics, auto instrumentation, focused metrics and ready-made dashboards. Core capabilities embody end-to-end distributed tracing, clever AI guardrails, value and token governance, complete stack monitoring, and compliance and auditing.
Boston-based Dynatrace initially launched AI Observability in 2024. Earlier this 12 months the corporate considerably up to date the product with an emphasis on autonomous operations and agentic ecosystems.
Dynatrace AI Observability helps greater than 20 main AI platforms together with OpenAI, Amazon Bedrock, Google Gemini, Google Vertex, Anthropic and LangChain.
JetStream Safety-First AI Governance Platform
JetStream Safety describes its “security-first” AI governance (SAIG) platform as a management airplane for AI.
JetStream supplies enterprises with visibility into how AI programs behave in manufacturing by way of its real-time governance, id, design management, and monetary accountability capabilities.
The platform is anchored by JetStream Blueprints, what the corporate calls “residing operational contracts,” that make each AI agent, mannequin, workflow and id inside an enterprise seen, attributable and ruled in actual time. It constantly discovers AI programs throughout the group, replaces uncooked credentials with digital revocable keys, enforces runtime coverage on each agentic motion, and ties workflows to accountable homeowners.
JetStream, a startup headquartered in Santa Clara, Calif., was based in 2025 by a number of former CrowdStrike and SentinelOne executives. In March of this 12 months the corporate exited stealth with $34 million in seed funding from Redpoint Ventures and numerous angel buyers together with CrowdStrike CEO George Kurtz, Wiz CEO Assaf Rappaport and Okta co-founder Frederic Kerrest.
LangChain LangSmith Platform
LangSmith is LangChain’s business platform for prototyping, debugging, testing, evaluating, deploying and monitoring giant language fashions, AI brokers and AI workflows constructed utilizing the open-source LangChain framework.
LangChain, based in 2023, develops the LangChain framework (obtainable beneath the MIT open-source license) for constructing AI functions and brokers utilizing LLMs, whereas additionally providing business merchandise reminiscent of LangSmith.
LangSmith’s core performance contains observability and tracing (trace-level debugging and multi-turn threads); analysis and testing (dataset administration and hybrid scoring); manufacturing monitoring (value and latency monitoring, error and suggestions loops); managed deployment (sturdy runtime and agent studio); and superior automation and agent runtime instruments (together with the LangSmith Engine and LangSmith Sandboxes).
LangSmith is framework-agnostic, that means that it may be used exterior 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 Sequence B funding spherical that boosted the corporate’s valuation to $1.25 billion.
New Relic AI Observability
New Relic AI Observability supplies end-to-end efficiency monitoring, value governance and real-time troubleshooting for AI functions and agentic workflows. It’s a core part of the corporate’s flagship New Relic Clever Observability Platform.
New Relic AI Observability’s capabilities embody AI stack monitoring and integrations (full-stack tracing, ecosystem connectivity, and mannequin efficiency mapping); agentic AI governance (multi-agent tracing, no-code agent builder, and Mannequin Context Protocol); value and danger administration (token monitoring, hallucination and drift detection, and enterprise guardrails); and generative troubleshooting and AIOps (pure language querying, root-cause reasoning, and log and anomaly evaluation).
In June the San Francisco-based firm launched a brand new characteristic known as New Relic AI Coding Observability that’s particularly designed for AI-assisted software program growth. It extends production-grade monitoring instantly into the coding part of the software program lifecycle, offering visibility right into a growth group’s use of AI coding assistants. The instrument helps builders measure productiveness beneficial properties from AI coding assistants and workouts value controls over their use.
Pendo Agent Analytics
Pendo Agent Analytics supplies a solution to measure the efficiency, utilization and enterprise affect of AI brokers throughout an enterprise. Constructed on the Pendo Software program Expertise Platform, Pendo Agent Analytics was launched to common availability in December 2025 following a six-month early entry trial interval.
Throughout that early trial, in keeping with Pendo, dozens of organizations used the software program to grasp how customers work together with AI brokers, measure the standard of agent outputs, and determine the place agent conduct supported or hindered person productiveness.
Core performance contains monitoring hybrid workflows throughout each brokers and conventional software program, surfacing patterns in how individuals work together with brokers, flagging “off-script” conduct reminiscent of hallucinations and lack of ability to reply, and mapping agent exercise to activity completion and general engagement for ROI and efficiency assessments.
In February, Raleigh, N.C.-based Pendo acquired Chisel Labs, a startup that develops an AI-powered platform that growth groups use to jot down product requirement paperwork and launch notes, conduct person analysis, and triage person suggestions.