Ask HN: Why are we humans still prompting to make Agents better?
A Hacker News user questions why humans still manually prompt AI agents instead of having agents self-improve, noting that many companies have raised significant capital for this goal but widespread autonomous agent improvement has not yet been productized or made generally accessible.
Background
- "Agents" refer to AI systems that can autonomously plan and execute multi-step tasks (e.g., booking a flight, writing and debugging code) without human step-by-step guidance — distinct from simple chatbots that answer one question at a time.
- The core idea: if an agent is smart enough to do useful work, why can't it also improve its own prompts, fix its own mistakes, or redesign itself — creating a self-improving loop? Many in the AI field call this "recursive self-improvement" or "agentic AI."
- Despite massive VC funding into companies building AI agents (e.g., Adept, Cognition AI, various YC startups), no widely available product has delivered on this promise. Current agents remain brittle, expensive per-use, and prone to failures that still require human oversight.
- The question reflects a known tension in AI: models like GPT-4 or Claude can outperform most humans on many narrow benchmarks, yet reliable autonomous agency — systems that truly "improve themselves" — remains unsolved, partly due to safety concerns, unpredictable behavior, and the lack of robust evaluation frameworks.