Velocity Through Rigor: The 30% Guardrail Rule for AI Teams

The proposal. Cate Huston, a former engineering leader at Automattic and DuckDuckGo, argues that the management practices that scaled human teams apply directly to AI agents. She proposes allocating 30% of total output to guardrails—tests, CI, process, architecture—based on her experience building two production products with a single engineer and Claude. The ratio is not arbitrary: her team's data shows that without this investment, incident rates and firefighting time eat into the gains from AI velocity.

The evidence. Huston cites Meta's internal data showing a 220% increase in changes but only 36% more user-facing impact, with a 40% increase in incidents and 70% more firefighting time. Faros AI and DORA 2025 confirm the pattern: throughput up, stability down. Her own team ships five releases per week with about 25 PRs each, but CI costs eventually overtook AI spend.

The trade-off. The catch is that investing in guardrails requires saying no to features. Huston's CEO asked only for features; she said yes to nothing for four months to re-platform off a no-code stack. The velocity win came from the platform change, not from AI alone. Leaders face the tension between visible feature output and invisible process work that prevents incidents. The 30% rule is a heuristic that forces the conversation.

What's missing. Huston's experience is from a two-person team using Claude on a single platform. Scaling to 20 engineers with multiple models and codebases introduces heterogeneity that the 30% rule may not cover. She acknowledges that medium-sized teams suffer most from variance. The video doesn't provide a framework for breaking down guardrail allocation across teams or measuring whether 30% is the right number for different contexts.

The managerial implication. Huston recasts the manager's job from scaling humans to scaling rigor in machines. The same layers—mission, strategy, tactics, process, execution—still apply, but each moves faster. She argues that process is how judgment scales into execution, and that agents multiply the variance of human habits. Engineering leaders must now budget for guardrails as first-class deliverables, not overhead.

The bottom line. The video's core actionable insight is that teams should measure guardrail investment as a proportion of output and target ~30%. The deeper lesson is that the management playbook is not obsolete: it's reconfigured for a world where execution is cheap but reliability is expensive. Huston's talk is a credible, experience-grounded argument for treating process as a product alongside features.