Shipping faster is compounding performance debt faster

Agentic coding is not just making developers faster; it is changing the risk profile of performance engineering. OpenAI reports a 70% jump in PRs merged per engineer per week, with nearly every engineer using Codex and every PR auto-reviewed. That volume shifts the balance between shipping features and protecting latency.

The non-obvious problem is compounding: every added check, network call, or data structure consumes a shared latency and infrastructure budget. Small regressions may be invisible per PR but accumulate faster when teams ship more. OpenAI says the bottleneck is not only GPUs; pre-inference work like auth, context fetching, tokenization, serialization, and memory is equally critical.

OpenAI's response is to automate performance engineering itself: agents triggered by regressions can profile, compare, propose fixes, and run benchmarks. The vision is always-on optimization agents, parallel skill sets, and autonomy loops. The evidence is partly aspirational; the talk offers directional data but few details on safety, false positives, or measurement validity.

For engineering leaders, the takeaway is governance, not speed: fast feedback loops, reliable benchmarks, observability, and test coverage become prerequisites for safe autonomous agents. Without them, teams get more changes, more hidden debt, and faster degradation. The question is whether your engineering maturity can survive the shipping-rate acceleration.