AI Turns Engineers into Planners and Reviewers
AI coding tools shrink writing time from ~4 hours/day to near zero, shifting effort to planning (saves 30min review per 5min upfront) and reviewing; parallelize agents past 5min executions to maximize throughput.
Prioritize Upfront Planning Over Iterative Reviews
AI tools like GitHub Copilot, early Cursor, and Claude Code have displaced most manual coding—previously ~4 hours/day—pushing engineers to spend that time planning tasks or reviewing AI outputs instead. Work doesn't vanish; it shifts, with only ~20 minutes regained per 30 minutes of prior coding after accounting for increased planning/review.
Two core approaches emerge: plan-heavy (write detailed specs, MD docs, or use interrogative prompting to exhaust edge cases) minimizes reviews by frontloading effort, yielding higher accuracy and fewer iterations. Review-heavy skips specs for quick starts (e.g., 'add contact form'), but demands constant context-switching fixes, wasting human time.
Always favor planning: 5 minutes upfront saves 30 minutes reviewing. Tailor by work type via this matrix:
| Feature Development | Migrations/Maintenance | |
|---|---|---|
| Frontend | Review-heavy (stateful edges like animations/styles hard to spec) | Plan-heavy (test-driven) |
| Backend | Plan-heavy (TDD feasible) | Plan-heavy (fully autonomous) |
Frontend features resist full specs due to interactions; backend/migrations suit hands-off execution.
Parallelize Agents to Handle 5+ Minute Runs
Agent capabilities grow— from seconds (Copilot line completion) to 30s (Cursor file), 1-2min (Claude Code last year), now 5-10min with tool-calling, type-checking, testing (e.g., Playwright MCP). Longer runs boost accuracy (testing > quick code gen), but cross the 5-minute threshold where staring at logs fails; humans multitask (Twitter) or parallelize.
Run multiple agents simultaneously: if each takes 10min, queue 3-4 so a fresh output awaits post-review. This maximizes human throughput as execution times hit 20min+ (forecast: AI soon QA's frontend via browser automation, slashing back-and-forth).
Build Interfaces for 'Focus Maxing' and Monetization Realities
Future tools must treat engineers as managers of parallel streams, not deep coders: enable task planning, QA assistance, AI/human code review, PR monitoring (auto-react to comments), previews/diffs in one view. Avoid 30s context switches that 'fry brains'; let agents run maximally before yielding control—'focus maxing'.
Vibe Kanban embodied this: sidebar for multi-agent workspaces (8 providers like Codex), Git diffs, inline comments, live previews. Launched June 2024, hit 30k MAU, 25k GitHub stars. Speaker demo'd live shutdown—AI added blog post, opened PR, deployed via Cloudflare—before announcing pivot to open-source only.
Shutdown rationale: mature market dominated by enterprise sales + token reselling (Vibe charged $30/mo but enabled $3k provider spends); no fun in '8th place'. Lessons: hire enterprise sales early, prioritize great teams/hard work (midnight Saturdays), build cutting-edge value (e.g., SweeBench leaderboard ahead of OpenAI). Next: time off, new ventures; regrets minimal.