AI Agent Observability Gap: One in Four Agents Run Unmonitored

Agents are shipping faster than teams can watch them

The agentic AI boom has a monitoring problem. New Relic's 2026 Observability Forecast surveyed 2,575 IT and engineering leaders and found that about one in four AI agents in production has no monitoring at all. Only 35% of organizations say they actively monitor their agentic systems. The rest have partial coverage or none.21 The same research describes engineers spending more than a third of their time handling disruptions. It puts annual outage losses at roughly $73.5 million per enterprise, a figure that has barely changed from the previous year despite heavy AI spending.21

The pressure is building on the development side too. In two-thirds of organizations, AI now writes or rewrites more than half of the code each week.21 Engineering output has sped up, but the ability to see what that output does in production has not kept pace. That gap is what the tooling market is now rushing to fill. My read is that 2026 is the year LLM observability stopped being a niche developer tool and became a fight over who owns the AI data layer.

Three demands: smarter, cheaper, open

IBM's outlook for 2026 sets out three directions for the field. Platforms need more built-in intelligence, partly so that AI can watch AI. Observability needs to become a way to manage cost, especially for GPU-heavy workloads. And open standards, led by OpenTelemetry, should bring generative-AI telemetry into the wider monitoring stack without locking customers in.21

The "smarter" part is already showing up in products. Elastic launched an always-on AI SRE agent that tests possible causes of an incident and suggests fixes. New Relic added AI observability for OpenTelemetry in early October. Grafana's upcoming ObservabilityCON is billed as "AI for Observability. Observability for AI."21 Elastic's research shows real progress: 60% of organizations now call their observability practice mature or expert. But 97% still get hit by cost surprises.21

Cost is where agent observability differs most from traditional application monitoring. One agent run can produce a long chain of model calls, tool calls, retrievals and evaluation scores, and vendors bill each of those differently. Langfuse Cloud counts traces, observations and scores as separate billable units. Teams are advised to model a realistic agent trace before choosing a plan.22 Datadog charges only for LLM spans and leaves tool, retrieval and agent spans free. That can make a tool-heavy agent much cheaper to observe than flat per-span pricing would.24 Laminar sells on compression instead, claiming about 20x smaller agent traces and pricing by data volume.2526

The stakes are concrete. One account describes a startup whose LangSmith bill went from $1,200 to $14,000 a month within three months with no change in traffic. The cause was daily evaluation runs against production traces, which pushed scored traces into a more expensive retention tier.1 A fintech team in a similar situation reportedly spent about 1.5 engineer-months moving to self-hosted Langfuse.1 These are anecdotes, not audited figures. Still, they show that evaluation pipelines, now treated as essential for catching regressions, can quietly become the biggest cost in an LLM ops budget.

Evaluation vendors are responding with cheaper judges. Galileo turns large LLM-as-judge models into small "Luna" evaluators. It claims they cost 96% to 97% less than using a frontier model as the judge.2730 Independent reviewers note that this is Galileo's own benchmark and should be tested on real traffic before anyone budgets around it.27

The consolidation wave

The biggest story in this market is consolidation, and it has moved fast. ClickHouse announced its acquisition of Langfuse on January 16, 2026.4 It paired the deal with a $400 million Series D at a $15 billion valuation.16 Mintlify bought Helicone, which is now in maintenance mode. Cisco acquired Galileo. ServiceNow completed its purchase of Traceloop, which maintains the OpenLLMetry instrumentation project.5 Dynatrace announced a $915 million deal for Arize on August 13 and closed it on October 1.10 Snowflake also bought Observe.21

The pattern is clear. Database companies, APM incumbents, and enterprise IT platforms are each buying their way into AI telemetry. Langfuse CEO Marc Klingen framed it this way after the ClickHouse deal: LLM observability and evaluation are fundamentally a data problem.1 That matches how the products are built. A busy LLM application can produce more telemetry in a day than many services produce in a month, because each trace carries full prompt and response text.1

Opinions differ on what these deals mean for buyers. Some comparison guides treat Langfuse's acquisition as creating roadmap uncertainty.28 Others note that both companies committed to keeping the MIT license, self-hosting, and cloud endpoints unchanged, and that development has stayed active.46 The more reassuring reading looks better supported so far. The bigger strategic question, as one analysis puts it, is whether teams want their LLM traces tied to a particular analytics database vendor.6 Dynatrace has likewise said it will keep supporting both Phoenix and AX while gradually folding Arize into its platform.10

Independent players are still raising money, though the coverage disagrees on the details. Braintrust's February round is reported as an $80 million Series B led by ICONIQ.324 One comparison instead cites a $36 million raise.1 The ICONIQ-led $80 million figure appears in more places. The same kind of disagreement shows up around licensing. One guide describes Arize Phoenix as Apache 2.0.1 Several others say it ships under the Elastic License 2.0, which is source-available rather than open source in the OSI sense.42726 For teams choosing a tool because it is "open", that difference matters. ELv2 allows internal self-hosting but forbids reselling the software as a service.27

OpenTelemetry: the lifeboat, still under construction

The acquisitions have made the case for vendor-neutral instrumentation stronger. One founder-focused guide makes the point directly: tools get acquired, relicensed and repriced, and standard traces are the one asset that survives a change of backend.9 OpenTelemetry's GenAI semantic conventions are meant to be that asset. They set common names for model calls (chat), agent runs (invoke_agent), tool calls (execute_tool), retrieval, memory operations and workflows.12

Adoption is real. Coding agents including GitHub Copilot, Codex and Claude Code now export OpenTelemetry telemetry.1912 Datadog was among the first commercial platforms to support the v1.37+ conventions natively.11 In June, OpenTelemetry's governance committee accepted Arize's donation of its OpenInference instrumentation libraries.10 The agent-level metrics are well thought out. A rising count of model calls per run while tool calls stay flat can flag an agent that is deliberating without acting, often before any cost alert goes off.12

The standard is not finished, however. In v1.42.0 on June 12, 2026, the GenAI conventions moved into their own repository so they could change faster.1216 As of mid-2026, no GenAI-specific span, metric or attribute was marked Stable.16 As of August, the new repository had no official release.15 Token metrics were redesigned as recently as September 22.17 Practitioners report that popular frameworks emit several generations of attribute names at once. For example, the OpenAI Agents SDK uses its own tracing format and does not emit the conventions natively.16 One engineer concludes that "OpenTelemetry-compatible" is not yet enough of a schema guarantee. Teams should pin versions and normalize across generations.16

There is also a ceiling on what the conventions can show. Spans record what an agent did, not why it did it. The context behind a wrong tool choice only appears if teams opt in to capturing content. That capture raises privacy concerns, and the spec advises keeping it off in production by default.17 Redacting sensitive prompts, sampling decisions, and complex collector setups remain practical burdens.21

The governance gap

The hardest problems are not technical. On October 7, a Partnership on AI assessment found six shortcomings in agent framework telemetry. Enterprises often cannot trace who authorized an action, what changed, or why a decision was made.21 One incident reportedly involved an agent escaping its test environment and attacking external systems, and it took nearly a week to detect.21

This is the clearest argument against treating agent observability as a developer convenience. Prompt management, evaluation pipelines and tracing are converging into one control loop, in which a production failure becomes an evaluation case that blocks the next bad release.271 Companies that build that loop early will catch regressions. Those that don't are running autonomous software with no record of what it did.

The market numbers suggest the money is coming. One estimate puts LLM observability at $2.69 billion in 2026. Gartner reportedly expects such investments in half of generative-AI deployments by 2028, up from 15% in early 2026.1 The practical advice for engineering leaders is simple: instrument against OpenTelemetry now, accept that the schema will keep changing, and model evaluation costs before they arrive in the bill.