Girish Konda

Girish Konda is a Seattle-based principal software engineer at Microsoft and the technical lead and architect for capacity management in Microsoft Fabric, where workloads from many customers share the same platform. Outside Microsoft, he built MCP Pacemaker so developers can restart an editor or exit a command-line session without shutting down the tool servers their AI agents use.

Girish studied computer science at Arizona State University and worked at Amazon and AWS before joining Microsoft. Before taking on capacity management, he worked on enterprise security for Fabric, including private network connections and encryption keys that customers control.

Finding the incident in the alert queue

When Girish and his colleagues tried using one agent to triage operational alerts, it could dismiss an alert just as confidently after recognizing its title as it could after checking the telemetry. They narrowed its job to grouping alerts by signature and routing them. A specialist for each kind of failure could then check the relevant telemetry against known benign patterns and escalate an unexplained alert to a person. The split also let the team test each specialist against past incidents. For Girish, the measure of success was how often the system caught a rare real incident in a queue full of familiar warnings. Girish’s account of redesigning alert triage

What survives a restart

In MCP Pacemaker, a long-lived local process runs the Model Context Protocol servers that an editor or command-line client would ordinarily launch itself. After a restart, the bridge can repeat a returning client’s opening handshake with a fresh server. A tool call needs different handling. If a tool creates a pull request and its reply gets lost, repeating the call could open a second pull request. Pacemaker leaves the outcome unresolved so the caller can check what happened.

Girish applies that reasoning to agents that delegate work and can change production systems. In his SAFE handoffs model, one coordinator checks the source and age of a worker’s evidence before acting on it. Each write gets its own identifier and single-use approval, so the coordinator can look up the effect when a reply goes missing. Girish makes a different tradeoff for starting servers. After simultaneous starts caused failures in a shared package cache, he limited how many could start at once, but let a waiting start proceed after a deadline. He accepted some risk to the cache to keep a developer’s request from hanging indefinitely.

In his October 2026 individual Internet-Draft on evaluation state, Girish asks what an engineer on call can learn from a refusal. A verifier might refuse a credential because it checked and found a revocation, or because the service that tracks revocations was unreachable. His proposal records which check the verifier could complete and what it knew at the time. That gives the engineer a reason to investigate the credential in the first case and the unavailable service in the second.