Nicole Forsgren on AI and Developer Productivity

A 32-minute interview on The Pragmatic Engineer, recorded 2026-05-18. Nicole Forsgren (formerly DORA lead, now at Google) discusses how AI has broken existing productivity measurement, why the bottleneck shifted from coding to everything downstream, and what actually matters for developer experience when agents always agree with you.

Summary

Key Points

1. AI speeds up coding (inner loop) but overwhelms everything downstream. Review, deployment, release, and policy processes that were built for human-paced throughput can't handle 10× code volume. The bottleneck shifted — "now we just threw gas on the fire."

2. Old organizational friction becomes a hard blocker. An intern committed more code before their laptop arrived than the system could process. Policies designed around weeks-long lead times break when new hires "can commit code on day one."

3. Faster feedback loops can increase cognitive load. AI may force you to rebuild your mental model "dozens of times in 30 minutes." Fast feedback used to be unambiguously good — now it's a double-edged sword.

4. Flow state depends on psychological safety, clear goals, autonomy, and purpose — not just tools. AI agents always agree with you, so they can't replace honest human feedback. Nicole threw away 100 pages of a book and relied on her co-author for the hard truth.

5. Measuring productivity requires asking "what do you mean by productive?" The SPACE framework (Satisfaction, Performance, Activity, Collaboration, Efficiency/flow) helps teams make intentional, risk-based trade-offs rather than blindly chasing velocity.

6. "Explicit executive sponsorship" — giving people explicit permission to experiment — reduces fear and creates psychological safety. People are afraid they'll be fired for making mistakes while experimenting with AI.

7. Burnout involves misaligned values, not just overwork. She cites Christine Maslach's research and recommends building a "personal board of directors" of trusted peers, often at other companies, who can tell you hard truths.

8. For agents or humans to self-improve, they need cheap, easy signals from the entire delivery pipeline. The outer loop needs to collapse the way the inner loop has. Start by identifying current touchpoints and quality gates now — knowing those signals will shift.

Key Quotes

• On rebranding: "If we can't call it developer experience, just call it agent experience, and then it's all going to work."
• On inner loop dopamine: "now we just threw gas on the fire."
• "Tech is easy, people are hard."
• On AI never pushing back: "I'm always so smart and I'm like, I know that was dumb."
• On human feedback: "our AI tools and agents just aren't there yet."
• On adoption metrics: "devs are like a gloriously cranky bunch of … We are not going to use tools that are awful"
• "If we just brute force it, something's going to break."
• On burnout: "Burnout is a combination of things… not having your values aligned."

Tools, Practices, and Methodologies

• SPACE framework — multi-dimensional productivity measurement (Satisfaction, Performance, Activity, Collaboration, Efficiency/flow)
• DevEx framework — flow state, cognitive load, feedback loops as the three pillars
• Adoption and engagement metrics — start here even if you hate them; developers won't use truly awful tools
• Personal board of directors — trusted peers (often at other companies) who give honest feedback
• Explicit executive sponsorship — telling people "you have permission to experiment" reduces fear
• Risk-based decision making — what you optimize for depends on what you mean by "productive"
• System instrumentation — agents and humans both need cheap signals from the full delivery pipeline

Unanswered Questions and Omissions

1. No concrete metrics for AI-driven productivity beyond adoption/engagement — the measurement gap remains
2. No resolution for two-person review regulations when agents write all the code
3. No systematic solution for maintaining mental models under AI-speed feedback
4. No exploration of over-reliance, skill atrophy, or AI-generated vulnerabilities
5. Governance implications of non-engineers building with AI left unexplored
6. No guidance for instrumenting legacy brownfield systems to surface the signals agents need

Transcript

Full transcript available at the source gist. 32 minutes 33 seconds, recorded 2026-05-18. Interviewer (Speaker A) and Nicole Forsgren (Speaker B).