ML engineer asks how to stop babysitting AI agents and get autonomous execution
Conscious-Shake391 · reddit · 2026-10-20
An ML engineer on Reddit vents about the core pain point of AI coding workflows: models aren't lacking intelligence, they lack autonomous thinking. He constantly has to ask the AI to rewrite its own prompts, remind it to research and compare alternatives, point out obvious flaws, and repeat instructions despite Custom Instructions and Memory. He wants: state a high-level goal → AI autonomously clarifies intent, researches, plans, delegates to subagents, executes, critiques, and delivers a verified result. The post poses four questions: is this a model reasoning or an agent orchestration/harness problem; are Hermes Agent, OpenClaw, Claude Code, or Codex actually solving it; can Skills, persistent memory, and automated evals reduce repeated corrections; and what's the simplest proven setup minimizing human intervention. A good gathering spot for power-user workflows.
Related event: ML Engineer Asks How to Stop Managing AI Like a Project Manager(2 posts)→
More from coding & agent
- 99% don't know 'record a skill': computer use agents' ceiling is nowhere near — alliekmiller · 2026-10-20
- Swapping to Qwen when GPT refuses your prompt is like typing sudo — DanGrover · 2026-10-20
- Letting an AI agent 'live a life' in a 30-year-old MMORPG: a Reddit experiment proposal — Aeylwar · 2026-10-20
- NYU professor builds EdgeBar with Cursor: a floating macOS sidebar for Edge profiles — kchonyc · 2026-10-20
- OpenAI Platform quietly adds a Decisions API playground with Predicate, Choice and Score modes — testingcatalog · 2026-10-20
- "Capable models don't imply productivity — they enable stupidity at superhuman scale" — generativist · 2026-10-20