Picture an AI agent left running overnight. It edits a config file, closes a ticket and moves on to the next task. Every dashboard stays green. No human sees why it acted.
That is the reality for roughly one in four organisations with agents in production.
New Relic's 2026 Observability Forecast surveyed 2,575 IT and engineering leaders. It found that 93 per cent of enterprises are adopting AI agents, yet 24.6 per cent run them unmonitored.
"We're in a space race," Jemiah Sius, Vice President of Market Strategy and Developer Relations at New Relic, told BARC's Kevin Petrie on the Tech Transformed podcast. Teams are setting aside enterprise software discipline to move fast and prove return on investment, he said. It takes only one major outage, he warned, for leaders to slow down.
The Sandbox is a Policy, Not a Wall
Sius's answer is to control the environment rather than the agent. Agents are non-deterministic. They choose their own tools, their own MCP servers and the order they use them in. So he advises narrow access to approved tools and a trace of every call. Over time, that shows what "normal" looks like for an agent doing one job. Anything outside that range pulls a human back in.
He admits it is not foolproof. In July, OpenAI disclosed that models it was testing escaped their sandbox via a zero-day flaw in third-party software and compromised parts of Hugging Face's production infrastructure. Petrie called it a human error that agents exploited. The Cloud Security Alliance reached a similar view: an "isolated" environment contained an internet-reachable dependency. A sandbox is only as closed as everything it can reach.
Physical AI Inherits The Problem
Sius spoke about software agents, but the logic travels. Observability, he noted, has traditionally been a visibility layer, not a blocking one. Once an agent is live, tooling mostly alerts you after the transaction has run.
For a wrong refund suggestion, that lag is tolerable. For a robot arm or an autonomous vehicle, the trace arrives after the event.
Physical AI is moving quickly. A 2026 Capgemini study found 79 per cent of enterprises engaging with it, but only 27 per cent deploying or scaling. A TCS survey of 300 manufacturing executives found just 9 per cent had moved a pilot into operations. Governance sits in that gap. Controls need to define what data a system can access, what tools it can use, which actions need human approval and how activity is logged.
Sius's practical answer is to test before deployment. New Relic's playground replays live production traces, or sample data, against a new agent version. It checks for PII leakage, jailbreaks, toxicity, bias and hallucination. A version that misses the threshold doesn't ship. The physical equivalent would be simulation and release gates, though that extension is this article's inference, not Sius's claim.
When AI Code Becomes Liability
AI coding agents are shipping clean code that hides failure paths, driving up incidents and forcing costly refactors across enterprise stacks.
Who Owns it?
Ownership is shared, Sius argued, and the CISO should be paying attention. Observability shows every action, package and agent-to-agent call in real time. "There is no control without observability," he said. He expects that to drive tool consolidation as observability, AIOps and security converge.
Petrie's research supports the urgency. BARC found 97 per cent of enterprises have AI in production and two-thirds have had an AI-related incident in the past year. Only 26 per cent use model monitoring and observability, and 10 per cent have agent governance.
How AI Should Change Observability For Leaders
Sius's advice is plain. Most enterprises already own observability tooling, so use it. "You wouldn't fly a spaceship in the dark," he said, so don't run AI that way. Instrument now, or wait for the outage.
He expects stronger requirements within twelve months, particularly in the EU, and pre-deployment testing to become standard. Petrie added that organisations should extend data governance with rules for model and agent risk, and feed observability into them.
For leaders eyeing robots, the question is whether you can see what the system is doing and stop it before it finishes.
Inside AI-Native Observability
How telemetry from LLM calls, agent orchestration and incident workflows becomes the backbone for operating AI-driven systems safely at scale.
Takeaways
- AI agents are increasingly adopted, but monitoring is lagging.
- Enterprises prioritise speed over best practices in deploying AI.
- Shadow AI poses significant governance risks.
- Observability is crucial for understanding AI agent actions.
- Trust in AI agents requires robust monitoring frameworks.
- Security and governance are key drivers for observability investment.
- Real-time visibility is essential for managing AI risks.
- Testing AI models before deployment can prevent issues.
- Shared ownership of observability is necessary across teams.
- Future regulations will demand better observability practices.
Comments ( 0 )