Everyone Bought AI Observability. Nobody Owns Agent Behavior.

94% of teams running AI agents in production have implemented some form of observability, according to LangChain's survey of 1,300+ practitioners. Yet 32% still name quality as the thing blocking them from shipping. That gap describes a category that sold everyone a solution to a problem it did not solve. The dashboards went in. The traces are flowing. The hard problem did not move.

I have watched this exact failure pattern before, and it did not resolve until someone named the gap correctly.

Observability records what happened, not whether it was right

A traditional application either works or throws an error. An agent completes successfully and returns the wrong answer. It calls the right tool with the wrong argument. It approves a transaction it approved differently ninety seconds earlier for identical inputs. It executes three steps of a five-step workflow, writes a partial state downstream, and exits clean.

None of that produces a stack trace. Latency is fine. Error rate is zero. The dashboard is green and the customer is furious.

The industry spent two years optimizing for hallucination because hallucination is the failure mode that makes headlines. ChatSee.ai's analysis of 10,000+ enterprise AI failure events across ten-plus industries found hallucination accounts for under 10% of them. The largest categories: resolution and escalation failures, decision failures, and execution and workflow failures.

We built tooling for the minority case and shipped the majority into production unmonitored.

Four markets, one label

Search "AI observability" and you will get comparison articles ranking Datadog against Langfuse against Zenity, which is like ranking a hospital against a pharmacy against a health insurer. They are not competitors. They serve different buyers with different budgets solving different failures.

Infrastructure telemetry. Datadog, Chronosphere, Dynatrace, and cost-restructured challengers like Oodle and groundcover. They answer "where in the stack did this break." Datadog reported $1.12 billion in Q2 2026 revenue, up 36% year over year, with an AI-native cohort above 750 customers, 31 of them spending over $1 million annually. Datadog also shows up below under autonomous remediation with Bits AI, an infrastructure incumbent extending into a newer segment rather than a pure-play competitor there. The incumbents are winning this, and they are winning it because when an agent times out you still need to know whether it was the model, the vector database, or a Kubernetes pod.

The developer inner loop. Langfuse, LangSmith, Braintrust, Arize, Opik, Pydantic Logfire. Trace, score, iterate, gate the release. This is where the open source energy lives. Opik processes over 40 million traces daily across 150,000-plus developers under Apache 2.0. At the time ClickHouse acquired it, Langfuse served 19 of the Fortune 50 and shipped 26 million SDK installs a month.

Runtime behavioral assurance. ChatSee, Zenity, Noma, WitnessAI, Cisco AI Defense, Palo Alto's Prisma AIRS. Did the agent behave within policy, and can you prove it to an auditor. This segment sells to security, not engineering, the same governance gap I covered in the shadow AI governance crisis spreading across Fortune 500 infrastructure.

Autonomous remediation. Resolve.ai, Traversal, Cleric, Datadog Bits AI. Observability data as input to an agent that fixes production. Resolve.ai raised $125 million at a $1 billion valuation in February 2026, months out of stealth.

Segment	Buyer	Example vendors
Infrastructure telemetry	Engineering / platform	Datadog, Chronosphere, Dynatrace
Developer inner loop	AI/ML engineering	Langfuse, LangSmith, Braintrust, Opik
Runtime behavioral assurance	Security / CISO	ChatSee, Zenity, Noma, WitnessAI
Autonomous remediation	SRE / operations	Resolve.ai, Traversal, Cleric

Four markets. Four buyers. One search term.

The acquisition map tells you what the buyers actually believe

Eight independent evaluation and observability companies were acquired in fourteen months. Weights & Biases went to CoreWeave for roughly $1.7 billion. Humanloop went to Anthropic. Statsig went to OpenAI for $1.1 billion. Langfuse went to ClickHouse. Promptfoo went to OpenAI. Helicone went to Mintlify. Galileo went to Cisco. Velvet went to Arize.

Run it alongside the infrastructure and security deals and the shape gets clearer. Palo Alto Networks bought Chronosphere for $3.35 billion, then Portkey, having already taken Protect AI and CyberArk. Snowflake bought Observe for a reported $1 billion. Cisco added Galileo to a security portfolio it has been building through acquisition for two years. SentinelOne took Prompt Security. Check Point took Lakera. Cato took Aim Security.

Three theses are being expressed with real money.

Foundation labs are buying the evaluation layer because whoever owns the testing surface makes the model underneath feel like a feature rather than a decision. Infrastructure vendors are buying the application layer above them because AI telemetry is becoming the most valuable workload on any database or compute platform. And security platforms are buying everything that touches agent traffic, because you cannot secure an agent you cannot see.

Notice which of those three is spending the most. Palo Alto alone has committed north of $3.35 billion on observability, and states its observability business has already crossed $300 million ARR. The security vendors are outbidding the AI-native vendors for the AI-native problem.

The bifurcation nobody is naming

Here is the thing that should worry you if you run an AI product.

Your engineering team bought Braintrust or Langfuse. They trace agent runs, score outputs, catch regressions in CI, and they are good at it. Your CISO bought Prisma AIRS or Cisco AI Defense or Zenity. They discover shadow agents, enforce policy at the gateway, and produce audit evidence. They are also good at it.

Neither team is monitoring the same thing. The engineering platform sees trajectories and cannot enforce. The security platform sees traffic and cannot reason about whether a decision was correct. And between them sits the actual question, the one that determines whether your agent is safe to run unattended: was this outcome correct and consistent given the input and the policy?

An OutSystems survey of 1,900 IT leaders found 96% of enterprises have AI agents in production and 12% say they can govern them. That 84-point gap is not a tooling shortage. Both teams bought tools. It is an ownership vacuum.

We have run this play before, with identity

When I was building LoginRadius, customer identity had exactly this shape. Product teams built login because they needed a signup flow. Security teams bought IAM (identity and access management) because they needed to control employee access. Both were correct. Both were funded. And for years nobody owned the customer identity surface, which is precisely where the breaches happened.

The category did not resolve when someone shipped a better login box. It resolved when CIAM (customer identity and access management, the consumer-facing counterpart to workforce IAM) got named as a distinct problem with a distinct owner, separate from workforce IAM and separate from application code. Naming it created the budget line. The budget line created the market.

Agent behavior is at the same stage right now. The tooling exists on both sides. The vocabulary does not. ChatSee is trying to name it as "runtime control," and Gartner is calling parts of it "guardian agents", which tells you the category has an analyst but not yet a consensus. Whoever names it convincingly will own the budget line for the next decade, the same way CIAM did.

What to actually do about it this quarter

Assume you need two tools, not one. A neutral evaluation platform for the engineering loop, plus either an APM (application performance monitoring) tool for stack correlation or a governance layer for audit evidence. Anyone selling you a single pane of glass across all four segments is selling you a roadmap.

Instrument to OpenTelemetry GenAI conventions anyway. With one caveat everyone gets wrong: every gen_ai.* attribute in the official registry still carries "Development" stability status as of mid-2026. It is not stable. It is still the only convergence path across LangGraph, CrewAI, and the OpenAI Agents SDK, so instrument against it and expect to migrate.

Make acquisition outcome a selection criterion. In this category that is not hypothetical. The larger deals preserved products. The smaller ones went into maintenance mode inside a quarter. Ask every vendor what happens to the roadmap if they get bought, and weight the answer.

Assign an owner for agent behavior before you assign a tool. If the answer to "who is accountable when an agent makes an inconsistent decision" is a tool name rather than a person, you have a governance gap that no purchase will close. A directory like GRC Compass is a reasonable starting point for mapping who else in the governance, risk, and compliance stack should own that seat.

Stop optimizing for hallucination. Look at your own incident history and classify by ChatSee's taxonomy: decision, interaction, workflow. If your failures cluster where theirs did, your evaluation suite is testing the wrong thing.

The prediction

By this time next year the standalone AI observability company will be rare. The infrastructure vendors will have absorbed the tracing layer, the model labs will have absorbed the evaluation layer, and the security platforms will have absorbed the runtime layer. What survives independently will be whatever sits in the ownership vacuum between them, because that is the only position none of the three acquirers can currently reach from where they are standing.

That is where I would build. It is also where I would look first if I were a CISO trying to explain to a board why an agent did something nobody authorized.

I am publishing the full market research behind this piece, including the segment-by-segment vendor breakdown. If you are evaluating a stack right now, start there.