Devs say LLMs are over-optimized for one-shot answers and refuse to ask for feedback mid-task
Elijah_Meeks · x · 2026-09-03
A developer complaint is resonating: models are poor at soliciting user feedback mid-task, which makes them worse at their job. The author has to painstakingly prompt LLMs to work in batches and hand over judgment lists at each step — and they comply reluctantly.
Elijah Meeks extends the point: models are severely over-optimized to one-shot everything (even via massive fanouts and a thousand turns) because it demos well to PMs and leadership. The thread argues agent UX is shaped by sales optics at the expense of genuine human-in-the-loop collaboration.
More from coding & agent
- Grok Bot at 16 days: a 7-step playbook for 24/7 multi-agent workflows — gekobraa · 2026-09-03
- Dev Configures In-App Purchases Entirely via ASC CLI + Agents, App Approved in a Week — rudrank · 2026-09-03
- Hermes Agent v0.21 'Pantheon' upgrade turns subagents into a self-coordinating team — gekobraa · 2026-09-03
- Constraining agents with LL(1) grammar + structured diagnostics: what it fixes and what slips through — Upstairs-Special-925 · 2026-09-03
- No-Code Dev Uses Claude, Meshy and Gemini to Build a Cozy Game — Gambo7592 · 2026-09-03
- Hemrock MCP: financial modeling prompts packaged as an MCP connector — modelcontextprotocol · 2026-09-03