Groundcover Closes the AI Observability Gap in Agentic Workflows
Cloud monitoring specialist Groundcover announced a significant expansion of its AI Observability platform on Wednesday. The update adds native support for agentic AI systems, full compatibility with Google Vertex AI, and a direct response to a problem quietly frustrating engineering teams everywhere. What actually happens inside multi-step agentic workflows.
The problem, at its core, is structural. Traditional observability tools target deterministic software systems where the same input reliably produces the same output. Agentic AI behaves differently. Dynamic prompts drive the outputs. Workflows stretch across dozens, sometimes hundreds, of model calls and tool invocations. As a result, the classic monitoring approach simply stops working.
Orr Benjamin, VP of Product at Groundcover, put the challenge plainly in a conversation with The New Stack. Moving from standard services to multi-step agentic workflows breaks the classic observability pillars of logs, metrics, and short traces. Instead of tracing a 20-hop microservices call, a team might need to reconstruct a two-hour session with 50,000 tool calls. The real challenge, therefore, becomes making sense of that data in a way engineers can actually act on.
To address this, Groundcover built its platform around a patented eBPF sensor that captures events at the kernel level. Because it operates below the application layer, it needs no instrumentation at all. Critically, this also means the platform sees what an agent actually did, not what the application chose to report. That distinction grows more important as agents become more autonomous. An agent that can manipulate its own logs or traces at the application layer makes application-level observability unreliable by design.
On top of that, the platform runs on a Bring Your Own Cloud model. All data stays inside the customer’s own infrastructure. Groundcover supports AWS, GCP, and Azure, and according to the company, this architecture cuts network costs by over 95% compared to traditional observability setups.
Since launching LLM Observability in August 2025, Groundcover has run across production AI environments throughout its customer base. This latest expansion delivers complete agent execution traces, covering every model call, every tool invocation with its arguments, and the reasoning path that connects them. Engineers can also adjust their focus level, moving from broad provider-level aggregates down to individual span detail depending on what the situation demands.
Token cost tracking has also grown more precise. The platform now tracks costs at the span level and accounts for the full pricing complexity of modern LLM APIs. Specifically, it distinguishes between regular input tokens, cache creation tokens, and cache read tokens. Consequently, teams can see the exact cost of each agent session rather than relying on rough estimates.
Benjamin offered a concrete example of why this cross-layer context defines strong AI observability for agentic workflows. A token usage spike that looks like an AI problem might actually trace back to a memory-starved Kubernetes node forcing retries. Because both infrastructure telemetry and LLM behavior live in the same platform, on the same timeline, and in the same query language, engineers can draw that connection immediately. That is the foundation Groundcover is building its case on for the multi-step agentic era.
