July 2026 Summaries

AI observability has advanced with the recent update to the MCP specification, which replaces custom logging with OpenTelemetry (OTel) data, enhancing traceability from tool calls through MCP servers to underlying infrastructure. This shift, rooted in recommendations from OpenTelemetry's GenAI Special Interest Group, allows observability tools to focus on their strengths without the need for custom MCP-specific tooling. Previously, MCP's protocol-specific logging isolated traces, hindering effective root cause analysis and requiring redundant efforts. The update enables full-stack observability, allowing a continuous trace through an MCP tool call, database queries, and infrastructure, thereby providing detailed insights into performance issues, such as identifying missing database indexes or resource constraints. This comprehensive tracing capability is crucial for ensuring AI agents adhere to predefined guardrails, as it allows engineers to verify that agents operate within expected parameters and offers a consolidated view of workflows, eliminating the need to manually integrate multiple observability tools. The convergence of OTel and MCP aligns with the broader ecosystem's move towards more effective and efficient observability practices, with companies like groundcover already leveraging these advancements to provide enhanced tracing capabilities.

Groundcover, an observability platform, has raised a $100 million Series C funding round led by One Peak, with participation from Morgan Stanley Expansion Capital and other investors, to advance its mission of building the world's first true observability platform for the AI era. The platform aims to provide AI-native companies with unsampled, high-fidelity telemetry data across infrastructure, applications, and AI workloads, using a bring-your-own-cloud (BYOC) architecture for privacy and control. Groundcover's approach addresses the limitations of legacy observability platforms, which often require data sampling and create gaps that hinder the performance of autonomous software agents. By offering full-fidelity telemetry and native OpenTelemetry support, Groundcover enables both engineers and AI agents to operate more effectively, reducing costs and improving data collection. The company has already attracted over 250 customers, including AI-native firms like TRM Labs and b.well, who have migrated from legacy tools to Groundcover's platform for better scalability and cost efficiency in managing their AI-driven operations.

Groundcover faced the challenge of implementing fast and scalable free-text search across vast datasets of logs and AI observability data, driven by the need for interactive searches that span hundreds of billions of log lines. The complexity arose from allowing customers to retain all relevant data, necessitating efficient query handling over large datasets. Groundcover experimented with various indexing strategies using ClickHouse, including native skip indexes, in-house inverted indexes for time-bound attributes, and eventually adopted ClickHouse's full-text inverted index for broader free-text search. They also employed a strategy of creating "favorite attributes" to manage JSON column path explosions, ensuring that only frequently queried fields were indexed. The solution involved optimizing query paths and indexing strategies to make searches efficient and cost-effective, ultimately enabling engineers to start investigations from familiar phrases or prompts.

Observability data is increasingly critical as it serves as the essential context for agents conducting system diagnostics and root cause analysis, making its protection and accessibility paramount. Traditional approaches that sample data or offload it to external vendors can lead to incomplete and inaccurate diagnostics, while methods like eBPF capture full-spectrum data at the kernel level, providing comprehensive insights without blind spots. Maintaining data residency within one's own cloud environment is crucial to ensuring security and cost efficiency, as demonstrated by groundcover's approach, which emphasizes that agents and models should operate within the owner's infrastructure. This strategy also involves careful management of external access through well-scoped connectors to avoid unwanted exposure, all while ensuring that agents can perform evidence-based diagnostics with complete and accurate data.