AI agent observability is not built in, and that is becoming its own line item

I read a launch this week that put a number on something I keep seeing in the wild. A company called Trust3 AI shipped a product on June 15 whose whole pitch is watching what AI agents actually do. In one of their customer stories, the tool caught an agent on track to burn through its entire monthly token budget in eleven days. Eleven days. Not flagged by the platform the agent ran on, not flagged by anyone on the team. Flagged by a separate tool bolted on top, after the fact.

That is the quiet shape of the year. Companies stood up agents fast. Then they looked up and realized they could not see what those agents were spending, touching, or deciding. And the place where that gap closes is turning into its own line on the budget.

The platform already watches the agents

Here is the belief I want to take apart, because it sounds completely reasonable. I bought an enterprise agent platform. It has dashboards. It shows traces. So observability is handled, it came in the box, no need to buy anything else.

I understand why that lands. Every major platform now ships with monitoring screens, and they look great in the demo.

The problem is what those screens were built for, and who they were built for.

The platform did build observability, for developers

The platform vendors are not lying to anyone. They did build observability. Developer observability, the kind an engineer uses to figure out why an agent got stuck in a loop, returned a bad answer, or blew past a latency target. That is genuinely useful, and if the job is debugging a workflow, it is exactly the right tool.

The trap is assuming that the tool an engineer uses to debug a model is the same tool a security lead uses to prove an agent never touched regulated data, or a finance lead uses to catch spend before it lands on the invoice. Those are different jobs, different audiences. The dashboard that satisfies the first one does almost nothing for the other two.

Most teams have not noticed the difference yet, because the bill has not arrived and the incident has not happened. That is the calm part of the story. The gap is fixable, and it is far cheaper to fix before either of those things shows up.

A separate control plane for agent oversight ships

Start with the launch itself, because the framing is the tell. Trust3 AI describes its product as a control plane that brings governance, security, and observability together, and the CEO, Balaji Ganesan, is blunt about who the existing tools left out.

I want to be fair. That quote is from a vendor selling exactly this, so take the marketing with the usual pinch of salt. But the underlying claim does not rest on trusting one company, because the rest of the market keeps saying the same thing from different directions.

Monte Carlo, the data reliability firm, shipped its own agent observability product back in March and split the problem into four parts: the context an agent relies on, its performance and cost, its behavior against the workflow it is meant to follow, and the quality of its outputs. The number reported alongside it by BigDATAwire is the one I keep coming back to. Seventy-three percent of enterprises say they will not ship an AI agent without monitoring and alerting, while 63.4 percent name the lack of monitoring and observability as a top barrier to deploying more.

Sit with that for a second. Nearly three quarters of companies refuse to ship without monitoring, and almost two thirds say missing monitoring is what holds them back. That is a market telling everyone the watching layer is not arriving for free.

Then there is Gartner, which in May predicted that only about 40 percent of organizations deploying AI will use dedicated AI observability to monitor model performance, bias, and outputs by 2028. Read it as a forecast and it sounds optimistic. Read it as a description of right now and it is sobering. If 40 percent is where the market is headed two years out, the majority of companies running agents today are flying with the cabin lights off.

One more signal, and this one is about money rather than surveys. Palo Alto Networks agreed to buy an observability company, Chronosphere, for 3.35 billion dollars, a deal framed around unifying observability and security for the AI era and closed in late January. When a security giant pays that kind of number to fold observability into its stack, the market is voting with its checkbook that watching agents is a security function, not a developer convenience that ships in the platform.

Agent observability is becoming its own category

So drop the idea that observability is a feature already owned. Pick up a more useful frame: observability for agents is becoming its own category, the way identity management and endpoint security did. It sits across whatever platforms a company runs, and it answers questions the platform dashboards were never designed to answer.

The question is not whether the agents are observable. It is whether the people who carry the risk, the security lead, the CFO, the auditor, can see what they need in time to act.

Held that way, the decision gets clearer. It is a build-versus-buy call, and a real one.

Building it is fair game. Plenty of capable teams wire up OpenTelemetry, route agent traces into the logging stack they already run, and stand up their own spend-and-scope dashboards. That works, and it avoids another vendor contract. The honest cost is ongoing. Advisory estimates floating around this year put the recurring overhead at a senior engineer plus sixty to eighty thousand dollars a year in observability and operations tooling, before anyone maintains it through every model and platform change. Build it when agents are core to the product and the muscle belongs in house.

Buying it means accepting a new line item in exchange for coverage that spans platforms on day one, with the security and compliance views already built. Buy it when the real exposure is regulated data, audit readiness, or spend nobody can currently see, and it needs to be visible this quarter rather than next year.

Neither answer is wrong. The wrong move is the third option, the one most companies are drifting into right now: assume it is handled, budget nothing, and find out otherwise when the token bill or the auditor arrives.

Can you answer what each agent spent and touched?

For a CEO or a Series C board with agents already in production, here is the version that fits on one slide.

We have agents in production. We can see what they do at a developer level, but we cannot yet answer, on demand, what each one spent last month, what data it touched, and whether it stayed in scope. Closing that gap is a deliberate choice between building it on our stack or buying a dedicated layer, and we are making that choice this quarter, before the next renewal, not after the first surprise.

That is it. No alarm bells. The companies that handle this well are not the ones with the most agents or the fanciest models. They are the ones who noticed, a quarter early, that watching the agents is a separate job from running them, and put a name and a budget against it.