Summary
Mike, co-founder and CTO of Thundra, presents at AI Engineer on an agentic workflow that automates continuous performance optimization for production applications. The talk opens with a familiar pattern: performance issues accumulate in silence, investigation timelines are unpredictable, and knowledge concentrates in individual engineers who eventually leave. Thundra’s solution is a weekly automated agent that analyzes real production telemetry, scores optimization opportunities by ROI, and opens pull requests with specific, grounded recommendations — turning the periodic performance sprint into a continuous background process.
The core technical contribution is what Mike calls “prod-to-code” context bridging: translating production metrics such as endpoint P90 latency, CPU, and memory into function-level context that coding agents can actually reason over. Without this translation, agents generate theoretically plausible fixes for code paths not exercised in production; with it, a seven-second endpoint can be traced to the specific function, database call, LLM invocation, or downstream microservice responsible. The talk references Google’s 2026 DORA metrics, which show that AI adoption has increased individual developer effectiveness while simultaneously raising software delivery instability — making automated quality and performance gates more important, not less.
Mike covers three recurring failure modes encountered in production: agents hallucinating fixes for inactive code paths, ClickHouse-specific SQL patterns confusing standard coding agents, and lazy exception-catching substituting for root cause analysis. The presentation is grounded in real deployment experience and offers a practical architecture for teams that want to move from reactive performance firefighting to proactive, agent-driven optimization.
📺 Source: AI Engineer · Published July 19, 2026
🏷️ Format: Deep Dive
