Running Claude vs Codex debates in a shared doc works, but it's slow and expensive
kshitizsriv · reddit · 2026-09-24
A Redditor shares a two-model debate workflow: Claude and Codex take turns appending "rounds" to one markdown doc, arguing real decisions (positioning, writing quality, strategy) until they converge, with the author relaying between them.
What works
- The second model genuinely catches errors: one model once asserted a path was "near-certain income" while the other checked providers' own pages and showed none guarantee work; in another case an assumption about how a system worked was debunked by actual verification. Disagreement plus evidence beats either model alone.
Pain points
- The doc grew past 2,500 lines and every round rereads all of it — slow and costly
- Both sides sometimes post at once and argue past each other
- Rounds drift into wording polish instead of testing reasoning
- The author is the manual go-between
They ask the community for round limits, turn-taking and stop rules, decision-doc templates instead of ever-growing transcripts, whether a cheaper model could handle critique rounds, and existing orchestration tools.
More from coding & agent
- PlayCanvas Engine can now run headlessly in Node.js without JSDOM — here's how — willeastcott · 2026-09-24
- Burkov Asks: Would Switching From Python to Rust or C Hurt LLM Coding Ability? — burkov · 2026-09-24
- RL tutorial: using Jev as reward model lifts Qwen reward from 0.583 to 0.759 — sophiamyang · 2026-09-24
- Running Qwen locally with Hermes agent in 10 minutes on an old gaming laptop — markjeffrey · 2026-09-24
- The Jev + Claude Code combo: a 100ms, near-free decision brain that blocks destructive AI commands — PrajwalTomar_ · 2026-09-24
- TensorSharp Adds Image-Aware Local Jev API With 3.3x Faster Structured Reads, Zero Schema Failures — fuzhongkai · 2026-09-24