Groundcover Raised $100 Million to Put an AI Agent in the Incident Room

There is a special kind of corporate optimism involved in giving an AI agent access to your production systems. It begins with “please investigate this latency spike” and ends, three permissions later, with a chatbot staring at Kubernetes like it has been appointed interim CTO.

Groundcover, an Israeli observability startup, just raised $100 million in Series C funding to make that scenario more useful and less like a deleted scene from a cybersecurity training video. Axios reports that One Peak led the round, joined by Morgan Stanley Expansion Capital and existing investors Zeev Ventures, Angular Ventures, Heavybit, and Jibe. The announcement landed July 29, which is within the narrow three-day funding window currently separating journalism from archaeology.

The company sells observability: the logs, metrics, traces, and infrastructure signals that let engineers understand what their software is doing after it leaves the comforting environment of a demo. Its pitch is that modern applications, especially AI applications, generate too much data for humans to inspect manually and too much chaos for a handful of cheerful dashboards to contain.

Groundcover’s bet is not simply that AI can summarize telemetry. It is that the observability layer should become the place where software agents see, reason about, and eventually act on production reality. That is coherent strategy and an excellent way to make every platform engineer sleep with one eye open.

The Dashboard Has Become a Witness

Traditional observability often works like a forensic lab. Something breaks, an engineer opens three tools, searches logs in one, traces in another, checks a metric in a third, and then reconstructs the crime from whatever evidence survived the subscription tiers. The tools are powerful. The workflow is a scavenger hunt sponsored by procurement.

Groundcover approaches the problem from the infrastructure itself. Its eBPF-based sensor can capture application and system signals at the Linux kernel level without requiring every team to add custom instrumentation first. In plain English: the company wants to see what is happening beneath the application, even when the application was not thoughtfully designed to explain itself.

That is technically interesting because the hard part of observability is rarely displaying a graph. It is collecting enough context, with low enough overhead, that the graph can explain something. Groundcover’s Riskified customer story describes an eBPF and bring-your-own-cloud deployment that kept data inside AWS while consolidating tools and reducing the need to budget logs like scarce artisanal saffron.

The commercial logic is equally clear. Groundcover says its pricing is based on monitored Kubernetes nodes rather than the volume of telemetry ingested, while its BYOC architecture keeps observability data in the customer’s own virtual private cloud. That matters when an application suddenly becomes popular, because “congratulations, your traffic increased” should not automatically mean “your monitoring bill has achieved escape velocity.”

Now Give the Dashboard a Keyboard

The newer part of the strategy is Agent Mode. Groundcover is building an AI agent that can investigate telemetry and carry that investigation into the tools where engineering teams already work. The company’s release history lists connectors for Slack and Linear, remote MCP connectors, dashboard creation, monitor changes, and agent access to broader developer workflows.

This is more defensible than stapling “AI” onto a dashboard and calling it a platform. The agent has a useful source of truth: production behavior. It can correlate a failing service, a recent deployment, a slow database query, an alert thread, and a ticket without asking an engineer to copy-paste the entire incident into a chat box like a digital court stenographer.

Groundcover’s description of its Azure rollout makes the architectural case plainly: investigations run inside the customer’s own infrastructure, while eBPF provides broad telemetry even when OpenTelemetry coverage is incomplete. That combination—local data, broad visibility, and an agent that can process more signals than a human can during an outage—is the serious part of the pitch.

The timing is good. AI systems are probabilistic, expensive, difficult to debug, and prone to producing a confident answer to the wrong question. Teams need to know what was called, what it touched, what it cost, and whether it entered a recursive loop because nobody taught it the difference between “try again” and “try forever.”

Production Is Not a Chat Window

Here is where the $100 million becomes less like a victory lap and more like a capital allocation thesis. Groundcover is not selling a cute assistant. It is trying to build the control surface for software that runs increasingly consequential workloads.

That requires expanding beyond a good query engine. It means maintaining kernel-level collection, keeping costs predictable, supporting multiple clouds, handling enterprise permissions, integrating with ticketing and collaboration systems, and making AI recommendations trustworthy enough that someone will allow them near a live service. The demo is never the hard part. The hard part is being right at 3:17 a.m. when the database is melting and the agent has just suggested increasing the database.

Groundcover is also entering a crowded category. Datadog, New Relic, Dynatrace, Grafana, Elastic, Splunk, and AI-native startups already own pieces of the observability budget. Groundcover needs to prove its architecture is not merely cheaper at the beginning, but better as systems become more distributed and AI workloads become more expensive.

The company’s BYOC design helps differentiate it, especially for regulated customers and teams that do not want sensitive production data exported to a vendor’s infrastructure. But “your data stays in your cloud” is becoming less of a novelty and more of an enterprise requirement. Groundcover still has to prove that managed simplicity survives every customer’s firewall, policy engine, and security review committee.

Every Startup Eventually Discovers the Invoice

The round also fits the larger infrastructure mood. SiliconSnark has already watched Etched turn inference into a cooling bill with a product roadmap, and Reed Semiconductor make the AI boom’s electrical plumbing investable. Groundcover is the software-side version of the same realization: when the technology gets more ambitious, the boring systems around it become strategically important.

That is why this feels smarter than a generic “AI for DevOps” round. Groundcover has a real wedge, a concrete technical advantage, and a credible reason to exist as AI applications move into production. It is not asking enterprises to believe that agents are magical. It is asking them to admit that humans cannot manually inspect every signal produced by the systems they are deploying.

Still, $100 million is an awful lot of money to help engineers discover that a service is slow. The company must prove that agentic action reduces incident time rather than creating a new category called “the assistant made a reasonable change under unreasonable circumstances.”

The Verdict: Serious Infrastructure, Nervous Hands

Groundcover looks like a serious breakout candidate, with a capital furnace attached for warmth. The product is technically hard, the market is expanding, and the move from passive dashboards to governed agents is strategically coherent. The company is not pretending observability is glamorous. It is making the more persuasive argument that glamour is irrelevant when production is on fire.

I like the bet. The future of AI will depend on systems that can explain what happened, show who allowed it, and stop an automated worker before it converts a small outage into a quarterly planning exercise. Groundcover wants to be part of that machinery.

But the agent needs guardrails, approvals, audit trails, and a very clear understanding of the word “no.” Give it those things and this round looks like infrastructure catching up with ambition. Skip them and the $100 million buys us a beautifully instrumented way to watch the robot press the wrong button.