Lore

CI Investment as the Velocity Lever

Overview

Under AI-heavy engineering workflows, velocity is bottlenecked by CI quality, not by agent capability. Agents can generate and iterate on code far faster than a human, so the limiting factor shifts to how quickly and reliably CI can validate what they produce.

Apply

When asked how to make an AI-augmented engineering team faster, invest in CI first — tighten tests, checks, and feedback speed — and treat weak CI as a reason to pause feature shipping rather than a reason to skip validation.

Related: Eval Investment Triage, Agent-Driven Exhaustive Experimentation (Practical vs. Theoretical Quality)

CI speed as a mathematical ceiling on agent throughput (Stripe example)

Ryan Nystrom argues CI/pipeline speed is not a minor developer-experience nicety but a hard mathematical limit on how much shippable value either humans or swarms of agents can produce: no matter how many agents you run in parallel, total throughput is capped by how fast changes can be built, tested, and merged. He cites Stripe's reported ~1,300 agent-authored PRs per week as evidence that without fast CI, high agent throughput is pointless — the PRs pile up faster than the pipeline can verify and land them. His conclusion: engineering leaders should allocate dedicated time to pipeline speed specifically because of AI, not despite it, since agentic engineering multiplies the cost of a slow pipeline rather than bypassing it.

Mathematical Ceiling on Throughput

Nystrom frames CI pipeline duration as a hard mathematical ceiling on ship rate, not just an inconvenience: "there is just a true mathematical limit on your capacity to ship code to production that is a reflection of how fast your CI pipeline is." He illustrates with a blunt contrast — a 1-hour CI loop versus a 3-minute one — arguing the difference caps how many iterations, human or agent, can complete per day regardless of how many agents run in parallel. His explicit recommendation: "allocate time to your pipeline, please," treating CI speed as a prerequisite investment before scaling up agentic engineering; more parallel agents against a slow pipeline, in his framing, just means "you might as well be throwing all those PRs in the trash."

He also ties CI speed to engineer psychology, independent of agents: slow CI pushes engineers to batch changes into large, cautiously-reviewed "monster changes," while fast CI supports a tight ship → feedback → iterate loop. He presents this loop as predating agentic workflows and merely amplified by them, not created by them.