AIエージェント実践録AI・テクノロジー
I Let an AI Agent Run for 48 Hours — Here's What Happened
Bottom line: humans are still necessary.
機械翻訳 / Machine-translated
2026年3月10日・4分で読める
Bottom line: humans are still necessary.
機械翻訳 / Machine-translated
I handed a task to an AI agent and let it run fully autonomously for 48 hours.
Bottom line: humans are still necessary.
But where humans are necessary has shifted completely from six months ago.
| Item | Setting |
|---|---|
| Model | Claude Opus 4 |
| Task | Add 3 API endpoints |
| Condition | Tests passing = done |
| Item | Rating | Comment |
|---|---|---|
| Code generation | ★★★★☆ | ~80% was usable quality |
| Test writing | ★★★★★ | Coverage exceeded 90% |
| Refactoring | ★★★★☆ | More consistent than humans |
| Item | Rating | Comment |
|---|---|---|
| Design decisions | ★★☆☆☆ | No explanation of "why this structure" |
| Edge cases | ★★☆☆☆ | Missed exceptions requiring domain knowledge |
| Prioritization | ★☆☆☆☆ | Treated everything with equal weight |
AI is a genius at "execution" and an amateur at "judgment."
In other words, human work has completely shifted to "deciding what to do." The time I spend writing code dropped to one-fifth, but the time I spend thinking about design doubled.
This isn't regression. It's simply a return to how things should have been all along.
Decisions are human work. The moment we delegate that to AI, humanity is finished.
This series will share every result from my AI agent experiments in full. Next time: a report on the "make AI earn ¥100,000 a month" challenge.
