When you’ve been going back and forth on a tricky problem, the agent is anchored — its context is full of earlier (possibly wrong) attempts. Hard to think fresh.
The trick: open a second session and ask the same question from scratch. No memory of dead ends. It often finds a simpler solution because it isn’t patching its previous approach.
Take it further with a writer/reviewer pattern: one session writes the code, a separate session reviews it. The reviewer catches things the writer normalized — subtle logic errors, missing edge cases.
Even further? Ask three separate sessions to each propose a different approach. Compare the three. This “competing hypotheses” pattern prevents anchoring and surfaces options you wouldn’t have considered.
This works with any tool — another terminal tab, a new Copilot Chat thread, or a separate Claude session.
💡 Try this: Next time you’re stuck, open a second session and re-ask the question without pasting in your previous conversation. See if you get a different (better?) approach.
🔗 Run multiple sessions in parallel (Claude Code best practices)
Do you ever use multiple agent sessions in parallel?
🟢 Yes — second opinions, parallel exploration
🟡 Occasionally, when I’m really stuck
🔴 No — one session at a time
⚪ Interesting idea, haven’t tried it
Have you tried the writer/reviewer pattern or something like it? How did it go?