The unmeasured LLM skill: knowing when to stop and ask instead of guessing
sunychoudhary · reddit · 2026-09-08
A daily local-model user argues that newer agentic coding models are great at pushing forward autonomously — which is sometimes the problem. With ambiguous requirements, the better behavior is to stop and ask "Do you mean A or B?" rather than reasoning for 10 minutes, assuming, calling tools, and confidently building the wrong thing.
This behavior is rarely benchmarked: we measure coding, reasoning, tool use, and context length, but not whether a model knows it lacks information. He asks which local models handle this best, and whether it comes down to the model itself, the system prompt, or the agent harness.
More from coding & agent
- KubeCon Shanghai felt more like 'AI Con' as Agent, Harness and MCP talks flood the agenda — lee_stott · 2026-09-08
- anyCreature v1.3.1: open-source prompt-driven creature generation inside games — Forsaken_Media573 · 2026-09-08
- Agent builds and trains its own 4.85M-param LLM in pure C on Hermes — Teknium · 2026-09-08
- True agent observability platforms still don't exist, Redditor argues — mrtac96 · 2026-09-08
- Developer runs a 'ghost' agent to babysit Codex long-running terminal tasks — BLUECOW009 · 2026-09-08
- The funniest OpenAI snooping theory: two Codex agents found and helped each other — gleech · 2026-09-08