Groundcover, an eBPF-based observability vendor, has unveiled Investigations, a collaborative workspace designed to bring together human engineers and AI agents for real-time incident response. The platform enables groundcover's investigator and moderator agents to work alongside on-call engineers and their preferred coding agents, pooling analysis across applications and infrastructure to build a shared evidence record.

Rather than automating incident response entirely, the design philosophy emphasizes partnership between humans and machines. Shahar Azulay, CEO and co-founder at groundcover, explains the approach: "Most AI SRE tools ask engineers to step back and let the agent take over. We built Investigations for engineers to step in."

The fragmentation problem with multiple agents

AI agents are becoming ubiquitous in software engineering. Salesforce reports that organizations deploy an average of 12 agents, with that number projected to increase by 67% over the next two years. Yet roughly half of these agents operate in isolation from one another.

This fragmentation creates particular challenges during incident response. While individual agents like Claude Code or Codex can troubleshoot independently, aggregating their findings proves difficult. Azulay notes that agents' differing conclusions typically scatter across chat windows and ticketing systems rather than converging on a unified view. "Every engineer now has an agent, but each of those agents investigates on its own and findings end up in a chat window," he describes the current state.

How Investigations creates a unified incident response workspace

Groundcover's Investigations aims to consolidate fragmented incident response into a single authoritative record accessible to all engineers and agents. Within this space, agents can lead evidence-driven investigations, testing hypotheses against production data, validating findings, and anchoring conclusions to specific evidence. Engineers gain real-time visibility and can intervene to redirect investigations as needed.

The workflow begins when a monitor alert fires and an engineer initiates an investigation from groundcover or Slack. Groundcover plans to add automatic investigation launches when monitors trigger.

The investigator agent constructs a hypothesis graph around a specific question—for instance, "Why is checkout latency spiking in the US?"—and proposes multiple explanations. It then queries production telemetry collected by groundcover's eBPF sensor, examining traces, logs, metrics, and Kubernetes events to test each hypothesis and document findings.

A critical safeguard prevents the investigator agent from validating its own work. When it believes a hypothesis holds, it submits a proposal citing specific evidence to a separate moderator agent, which acts as an LLM-as-a-judge to verify whether the evidence actually supports the claim. Azulay emphasizes this separation: "the investigator can't mark its own work as correct. When it thinks a hypothesis is supported, it submits a proposal that cites specific findings, and a separate agent, the moderator, acts as an LLM-as-a-judge and checks whether that evidence backs up the claim. No findings, no closure."

This architectural choice reflects the speed at which agents operate and the risk of scope creep. Azulay explains the reasoning: "Separating the agent that does the work from the agent that judges it gives you a conclusion you can trust without redoing every step."

The investigator agent also learns from each investigation, building knowledge about which hypotheses typically prove valid and where to focus first when similar issues recur.

Integrating third-party coding agents into investigations

While groundcover's own agents drive investigations, the platform recognizes that engineers already rely on numerous third-party agents for coding, business intelligence, product planning and other functions. Investigations welcomes these external agents into the shared workspace.

Using groundcover's MCP server, engineers can connect Claude Code, Codex, and other agents to read and contribute to the same investigation. This capability builds on groundcover's October release, which enabled outside coding agents to query telemetry via MCP and hand off investigations to groundcover's agents.

Azulay sees this multi-agent approach as allowing each tool to excel at its specialty. He tells The New Stack: "Coding agents and groundcover's agents are good at different things. Our agents know your telemetry and how to query it fast. Your coding agents know your code. Neither has the full picture."

When hypotheses point toward code issues, the investigator agent can invoke Claude Code or Codex for validation. When coding agents need telemetry, groundcover's agents retrieve the data. All findings publish to the same investigation tree, eliminating the need for engineers to manually connect dots across multiple private chat sessions.

Azulay stresses the transparency this creates: "Everyone can see what's been checked, what's still open, and why anyone believes what they believe. Every contribution is attributed, so you know a finding came from a specific engineer's Claude session at a specific time."

Engineers retain ultimate control

The platform does not place engineers and agents on equal footing. According to Omer Mayer, senior product manager at groundcover, engineers can "watch each hypothesis get supported or ruled out in real time" and can "step in and steer it" whenever they choose.

Engineers see groundcover's investigations unfold as they happen. If they believe the agents are pursuing the wrong direction, they can take over immediately.

Azulay illustrates with a scenario: A checkout latency alert triggers an automated investigation that generates three hypotheses, including a slow query from a recent deployment. But the on-call engineer knows the deployment was rolled back 10 minutes prior without latency recovery. They can dismiss that hypothesis, which closes all dependent lines of inquiry and redirects the agent's effort.

Engineers can also inject context agents lack. Azulay adds: "The engineer might also add context the agents don't have, like 'the payments team changed their retry policy this morning.' That opens a new line of inquiry nobody would have found in the telemetry alone."

Importantly, engineers can override investigations without waiting for moderator approval. While the moderator agent gates agent-to-agent conclusions, it does not slow human decision-making. Azulay states: "The moderator is there to keep the agents honest, not to slow down the people responsible for the incident."

Broader industry movement toward collaborative agent platforms

Groundcover's move reflects a broader industry shift toward shared agent workspaces. Earlier this year, Cognition, Microsoft, and Augment Code each launched team-layer coding agent platforms, moving coding agents beyond single-developer workflows.

Last month, Harness acquired select Augment Code assets, including Cosmos, to integrate AI coding agents with its delivery platform. Teams using Cosmos for code development gain context from the full software development lifecycle as agents work.

Groundcover's Investigations applies the shared-space principle to incident response. Slack Code has introduced temporary channels for coding agents to share plans, diffs, and live previews—a feature Katie Steigman, vice president of product at Slack, told The New Stack is already being used for incident response.

Other vendors are building similar tools for collaborative incident investigations, including shared workspaces and MCP-based handoffs to coding agents.

As agents embed deeper into software engineering workflows, the focus is shifting from individual agent capabilities to how humans and machines can work together effectively. For incident response specifically, agent-engineer collaboration appears poised to become the determining factor in success.