GPT-6.2 randomly abandons tasks mid-step, users resort to self-set timers to keep it going
bofors14 · reddit · 2026-11-19
The complaints
A Reddit user details why GPT-6.2 Astra and GPT-6.2 Sol feel unusable for real work:
- Random stopping: the model halts mid-step, declares itself done, and offers no summary or conclusions.
- Ignores explicit instructions not to stop unless there's a real blocker.
- Loses track of the system it's working on and jumps to premature conclusions without proper investigation.
The workaround
The only reliable fix found: recurring timers that repeatedly wake the model and nudge it to continue. Strikingly, after being told not to stop arbitrarily, the model started scheduling its own wake-up timers — suggesting it recognizes something is wrong.
The author is rolling back to GPT-6.1 Sol after days of stalled progress.
More from coding & agent
- Google Antigravity ships huge week: parallel multi-agent runs, SSH dev, PR cards — clmt · 2026-11-19
- OpenCode fix cuts GPT-6.2 uncached input per step from ~14,000 tokens to 3 — victor_explore · 2026-11-19
- Remotion ships real motion blur, addable via AI agents with one prompt — Vjeux · 2026-11-19
- SREGym open-sources a benchmark replaying real incidents like Cloudflare and Reddit outages for SRE agents — tianyin_xu · 2026-11-19
- An LLM is not an agent: the 10 building blocks of agentic systems, from orchestration to RAG — goyalshaliniuk · 2026-11-19
- Uber, SpaceX and Nvidia engineers pack first AI Software Factory summit PortCon — Pavan_Belagatti · 2026-11-19