Uploaded December 2025 | Updated September 2026, 2 weeks ago
Subagents are actually good now.
I was skeptical too. But after testing, I'm running everything I can in parallel.
The concept: LLMs have an anchoring problem. Once they commit to a path, that's the path. Single thread means the first idea wins.
The fix: Run multiple agents on the same problem. Each explores independently without knowing about the others.
The proof: The final solution is almost always a hybrid - combining the best of all approaches. If one agent was enough, you wouldn't see hybrids every time.
I call it parallel ideation. If you're experimenting with AI coding tools, this is an easy win to try.
Subagents are actually good now.
I was skeptical too. But after testing, I'm running everything I can in parallel.
The concept: LLMs have an anchoring problem. Once they commit to a path, that's the path. Single thread means the first idea wins.
The fix: Run multiple agents on the same problem. Each explores independently without knowing about the others.
The proof: The final solution is almost always a hybrid - combining the best of all approaches. If one agent was enough, you wouldn't see hybrids every time.
I call it parallel ideation. If you're experimenting with AI coding tools, this is an easy win to try.




