GPT-6 early access review: multi-session collaboration, pruning AGENTS.md made it 100x better
patricksrail · x · 2026-09-04
An early-access GPT-6 user shares hands-on findings from extended use in Codex Desktop:
- Heavy multi-session collaboration: complex projects benefit from a "babysitter" session keeping others on track; inter-model messages evolve into a kind of compressed English.
- Long-running tasks: works better without goal mode; the author had tasks run for days and monitored progress via an auto-updated HTML file.
- Prune your AGENTS.md: removing all instructions about subagents, docs, and testing made it "100x better"—keep only environment-specific setup.
- Strong self-validation, including UX testing; only weak spot was troubleshooting slow DB queries.
- Frontier-AI speak traces (hyphenation, word concatenation) can be prompted away permanently.
- Counterintuitively, one-sentence "caveman" prompts outperformed detailed ones.
More from coding & agent
- WHALE paper: alternating weight and harness optimization lifts agent accuracy by up to 24 points — Kangwook_Lee · 2026-09-04
- Your AI assistant only stores, never updates: a dev builds human-like memory with VectorAI — PrajwalTomar_ · 2026-09-04
- AI Won't Kill the Law Firm Apprenticeship Model—It Can Fix It — hoofnagle · 2026-09-04
- OUI-1: a fine-tuned Diffusion Gemma for Generative UI, 8x fewer params, open weights — GlennCameronjr · 2026-09-04
- GPT-6 Astra debuts at No.1 on Terminal-Bench, 1.9% ahead of Claude Fable 5.1 — sandersted · 2026-09-04
- Perplexity API lands in Stripe Projects: one CLI command provisions key and credits — jeff_weinstein · 2026-09-04