Groundcover: Observability Built for Agentic AI Workloads, Not Just Microservices

Classic observability was built for deterministic systems: a request comes in, spans emit, latency is measured. Agentic AI workloads don’t behave that way — they branch, retry, call tools, accumulate token context across multiple LLM calls, and fail in ways that don’t map to HTTP error codes. Groundcover’s VP of Product Orr Benjamin put it directly: the observability model that works for microservices is breaking down for multi-step agents.

Context

Platform engineering teams building agentic AI in production are discovering a category gap. Their existing observability tooling — Datadog, Grafana, OpenTelemetry pipelines — was designed for request/response workloads. It can emit spans from LLM calls if instrumented correctly, but it has no native concept of agent sessions, tool orchestration chains, or the relationship between prompt caching behavior and actual token costs. The result is teams flying blind on the metrics that actually determine whether an AI feature is working and what it’s costing.

Groundcover’s eBPF architecture is relevant here because kernel-level telemetry collection doesn’t require code instrumentation — it captures network calls, process interactions, and system-level events automatically. That matters for agentic workloads where the agent’s tool calls and external API interactions are part of what needs observing, not just the application code calling the LLM.

The Detail

Groundcover is extending its eBPF-based platform with capabilities specifically designed for agentic workloads. The most operationally significant addition is token cost tracking at the span level with prompt caching awareness — the platform can attribute token costs to specific agent steps, accounting for the discount applied when cached prompt prefixes are reused. For teams running long-context agents with persistent session state, this is the difference between accurate cost accounting and rough estimates.

The platform operates as BYOC (Bring Your Own Cloud) — deployed inside the customer’s AWS, GCP, or Azure environment rather than sending telemetry to a SaaS endpoint. Groundcover claims 95% network cost reduction compared to egress-heavy SaaS observability. For AI workloads generating high-volume trace data across multi-step agent sessions, the egress cost difference is material.

The multi-step agent tracing capability traces execution chains across tool calls, sub-agent invocations, and retrieval operations — giving platform teams visibility into where agent sessions are spending time and failing. Current observability tooling can capture individual LLM API calls; Groundcover’s model connects those calls into the agent session context that makes them meaningful.

Source: Groundcover Blog

Content created with AI assistance and reviewed for accuracy.