A revenue drop might come from a failed rollout, a local holiday, or a competitor's promotion. Investigating it means checking technical signals alongside the business context.
Leonel Candido da Silva, Senior Engineering Manager at eBay, joins us in the studio at Agent Conf 2026 in Warsaw to explain how his team uses AI agents in its Business Reliability Platform. We discuss specialized agents connected through MCP, a supervisor that prepares reports for the on-call engineer, and plans to use historical incidents to inform future investigations.
Checking business context alongside technical signals
Leonel's team monitors how business metrics respond to seasonality, user behavior, and internal failures. He describes introducing agentic AI when deterministic approaches were no longer sufficient to reduce noise or investigate the signals.
The agents check possible causes across systems, including Prometheus metrics, experiments, and application rollouts. They can drill down by location or platform and check whether a local holiday could explain a change. Leonel also gives Amazon Prime Day as an example of an external event that can affect revenue.
Specialized agents, MCP, and a supervisor
Within this workflow, Leonel describes about eight main systems, each with an agent specialized in accessing its data through MCP, the Model Context Protocol.
The agents process the data before passing information to a supervisor agent. This reduces the context the supervisor needs to handle. The supervisor prepares a report that goes into the PagerDuty description, giving the on-call engineer an initial triage to work from.
Giving the on-call engineer evidence and next steps
The agents suggest possible causes and next steps for further investigation. Engineers then verify the findings and decide whether to escalate or wait to see whether the data catches up.
Leonel distinguishes external influences on sales from internal problems such as delayed data or machines going down in a data center. Understanding which explanation fits the alert guides the investigation and the decision about how to respond.
Alert-triggered investigations and plans for earlier detection
Leonel describes monitoring real-time events, with an alert triggering the investigation. When asked whether the system is predictive, he says that is not the goal yet.
He explains that processing events such as clicks, impressions, and sales can take hours. In the next phase, he wants to identify problems upstream before their effects appear in business metrics, using that processing window to investigate and address them earlier.
Learning from historical incidents and investigation conversations
The team plans to bring historical cases into its agent workflow and correlate new problems with earlier incidents. Leonel describes an existing collection of root cause analyses in wiki pages and documents that is not yet standardized.
He also explains how the team uses agents to summarize Slack conversations during incidents and incorporate information from those investigations into their context. Internal documentation and accumulated business knowledge are further sources for the agents.
Coding, testing, and discovering internal data
Beyond monitoring, Leonel describes using agents to create pull requests that fix problems, increase test coverage, and test code and user interfaces.
The team is also working on discovering data across existing internal catalogs. Leonel explains that creating a centralized catalog would require migrations, coordination, and standardization that could take months or years. They are exploring agents as a way to consolidate information across those catalogs without first completing that migration work.
About the guest
Leonel Candido da Silva is a Senior Engineering Manager at eBay, where he leads a team working on the Business Reliability Platform.
At Agent Conf 2026, he presented “Agentic Monitoring in Production: How AI Agents Reduced Alert Investigation Time by 70%” together with Sujith N S. This studio interview follows up on the team's agentic monitoring architecture, the engineer's role in investigating alerts, and the next phase of work with historical incident data.
Resources
