- The author's AI token usage increased from $600/month to $25K/month after joining Lovable, with an overall spend of $85K since January.
- Development evolved from a solo dev with few agents to a system with a dedicated agent managing swarms of subagents, boosting merged PRs from 20-30 to 150+ per week.
- Human review is now reserved for high-impact changes like RFCs and ADRs, not traditional code review, shifting focus to system design and infrastructure decisions.
- AI workflows classify PRs by risk level (high, medium, low), enforcing human review for high-risk changes and using AI reviews for others to manage scaling and risk.
- Different AI review tools were tested: CI reviews serve as a safety net, while local reviews with subagents handle bulk work, though sharing setups between people remains unsolved.
- PRs remain useful as an abstraction for unverified changes, and stacked PRs help manage context and risk, as large PRs are not reviewable by AI or humans effectively.
- Managing human and agent attention involves using fast mode for interactive tasks, ensuring task autonomy to minimize permission requests, and batching validation work per PR stack.
- Agents require external task trackers like Beads, subagents for large tasks, small PRs to avoid context limits, and new context per task for basic hygiene.
- Task tracking for humans and agents should be separate to avoid clutter from agents' ephemeral text, keeping systems like Linear clean.
- Improving the development process involves agents writing skills, reflecting on past work, sharing context via git, and encouraging cultural sharing, though sharing skills is still challenging.
- The author merged 293 PRs in a week with no production defects, viewing the productivity gains as beyond unassisted development, despite token costs, with focus on outcomes over efficiency.
- The journey is described as fun, with anticipation for future innovations, and Lovable is hiring for those interested in this approach.