Groundcover calls observability OS

Observability started as software telemetry: logs, metrics, and traces that help engineers reconstruct what happened after a service fails. In AI systems, the same data now has to explain what an agent saw, chose, and executed step by step. (groundcover.com) (snowflake.com) groundcover Chief Executive Shahar Azulay told TFiR that observability is becoming the “operating system” for AI-driven software development, because agents need live production context to build, test, and debug code autonomously. (tfir.io) groundcover’s current AI observability product already reflects that argument. Its docs say eBPF auto-detects supported AI API calls with zero code changes, while OpenTelemetry instrumentation adds trace trees, tool execution chains, and multi-step agent workflows. (groundcover.com) The company says eBPF captures model, tokens, cost, latency, and full prompt and response content for supported providers including OpenAI, Anthropic, and Amazon Bedrock. It also says every generative AI span is stored by default, with no sampling or dropping. (groundcover.com) That is a different job from classic observability dashboards built for request errors and infrastructure saturation. Agent systems can fail by choosing the wrong tool, leaking data into a prompt, looping through a workflow, or taking an allowed action at the wrong time. (tfir.io) (snowflake.com) Other vendors are building adjacent pieces of the same stack. AgentField describes itself as an open-source control plane where every agent call is routed, every action is signed, and runtime policies and role-based restrictions are enforced during execution. (agentfield.ai 1) (agentfield.ai 2) Runloop is pushing the evaluation side. The company launched Benchmark Job Orchestration on April 24 with a Weights & Biases integration, saying teams can run agent benchmarks across thousands of environments and inspect trace-level behavior instead of only final scores. (prnewswire.com) (runloop.ai) Snowflake is tightening access around the same kind of data. Its Cortex Agent monitoring docs say teams can inspect conversation history, planning traces, tool selection, execution results, and final responses, while a new account privilege controls who can read unredacted observability fields. (snowflake.com 1) (snowflake.com 2) (snowflake.cn) The new Snowflake privilege is named `READ UNREDACTED AI OBSERVABILITY EVENTS TABLE`, and Snowflake says it is off by default for all roles. Roles without it can still read metadata such as tool names, token usage, latency, model name, and error severity. (snowflake.cn) (snowflake.com) The common thread is that traces are no longer just forensic records. They are becoming the place where teams decide what an agent is allowed to do, what it actually did, and who gets to inspect the raw evidence afterward. (tfir.io) (agentfield.ai) (snowflake.com)