Reddit practitioner: turn off thinking mode for agents, add a supervisor to break deadlocks
ngcheck03 · reddit · 2026-11-14
The author argues most agent work is tool calling that doesn't need thinking: it just generates tokens before answering, slows things down and bloats context, so he only picks models where thinking can be switched off (e.g. DeepSeek).
DeepSeek however tends to fall into micro-patching and deadlocks — a post-training quality issue that gets worse with thinking on. His fix: add a supervisor/advisor that checks the trace every 7-8 tool calls. Because the supervisor has a different perspective (even with the same model), it helps the agent escape rabbit holes. He admits this depends heavily on model quality and speculates strong models like Claude Opus may not need it.
More from coding & agent
- Free open-source "AI Engineering from Scratch" course teaches algorithms from raw math — ghumare64 · 2026-11-14
- Hugging Face Introduces Decision Models That Output Probability Distributions Without Generating Text — ariG23498 · 2026-11-14
- Inside the MI300X: 23-min visual guide from compute chiplets to a Gluon GEMM — saurabhtwq · 2026-11-14
- Job hunter's trick: fix a real issue in a startup repo, send the PR as evidence — minhnguyen · 2026-11-14
- Screenshot shows multi-agent, multi-user chat with realtime 3D video and long-term memory — nptacek · 2026-11-14
- DeepSeek opens its entire app stack via API: search, voice, OCR, ASR, TTS — churchkey · 2026-11-14