May Walter describes a workflow for finding performance problems that engineering teams otherwise struggle to prioritize. Production traces, query behavior and latency are connected to individual functions so agents investigate observed problems instead of offering plausible but unverified static-code guesses. The presentation is a Hud case study, not independent certification of its results.
The workflow combines scheduled GitHub Actions, a coding agent, MCP-accessible runtime information and skills for repeatable investigation. Suggested fixes are tested against the affected behavior before being surfaced for review. May Walter contrasts this approach with merely suppressing exceptions and explains why both excess low-signal context and missing forensic evidence weaken an agent's conclusions.
Human attention remains part of the system. Rather than flooding a team with pull requests, the workflow ranks opportunities by frequency, business importance, expected benefit and change risk, then presents concise evidence. May Walter argues that dependable autonomous workflows require stronger verification and operational guardrails than interactive coding assistance. Closing invitations are omitted.
Watch on YouTube
