🤝 AI Engineering Productivity is Anything But Normal

🤝 AI Engineering Productivity is Anything But Normal
This July 21, 2026 piece (cross-posted from Tom Tunguz's blog on Theory Ventures) argues AI-coding productivity gains cluster into three tiers rather than a single average. Tier one — simply distributing an AI IDE with no process change — yields roughly 20-46% gains (a Google RCT found ~21%, a GitHub/Microsoft/Accenture study ~24%). Tier two, the 'frontier,' reaches 2.5-3x when companies build an orchestration layer integrating tools like GitHub, Linear, and Slack, citing NVIDIA (3x code committed across 30,000 developers), Anthropic (2.5x code per engineer via internal Claude Code use), and Replit (tripled output). Tier three, the 'software factory' level, hits 8x or more when agents act as autonomous organizational units — Nubank achieved 8x efficiency using Cognition's Devin, and Goldman Sachs is piloting Devin alongside 12,000 developers. The key takeaway for fintech and enterprise-software investors: the gap between tiers comes from operating discipline and harness design, not underlying model quality.
🔗 Read more