Agents Are Dumb, Humans Must Tell Them Everything
Agents Are Dumb, Humans Must Tell Them Everything
“Machines don’t have common sense – we have to give them every detail, otherwise they hallucinate.”
In the era of ever‑more capable language models and autonomous agents, one uncomfortable truth persists: agents have no common sense. They can parse code, synthesize text, and even run tools, but they lack the intuitive world model that humans take for granted.
Why Agents Appear “Dumb”
| Reason | Explanation |
|---|---|
| Statistical Nature | LLMs predict the next token based on patterns in training data, not on an understanding of reality. |
| Lack of Embodied Experience | They never touched a cup of coffee, never felt rain, and therefore cannot infer physical constraints without being told. |
| No Persistent Memory of Real‑World Rules | Unlike a human, an agent does not retain a mental model of gravity, ethics, or social etiquette unless explicitly encoded. |
| Prompt‑Dependent Behavior | Small changes in wording can cause wildly different outputs – a sign that the agent is following surface cues rather than deep reasoning. |
The Human‑In‑The‑Loop Imperative
- Explicit Specification – Every assumption (file paths, network access, UI expectations) must be spelled out in the prompt.
- Verification & Guardrails – After an agent produces a plan, a human must verify its feasibility before execution.
- Iterative Feedback – Agents improve only when we correct their mistakes; they do not self‑correct based on common‑sense intuition.
Practical Tips for Working with Agents
- Use Structured Prompts – Break tasks into numbered steps; include pre‑conditions and post‑conditions.
- Add Safety Checks – Wrap any file‑system or network operation in a confirmation step.
- Leverage Tooling – Use
run_commandorgrep_searchonly after the agent has been instructed to do so. - Document Assumptions – Keep a markdown “Agent Playbook” that lists known limitations and required explicit instructions.
A Humorous Perspective
If an agent were a teenager, it would be the one who asks, “Do I really need to know why the sky is blue, or can I just copy‑paste the answer?”
That’s exactly why we must teach the agent, not expect it to know. The more we codify our expectations, the more reliable the output.
Looking Forward
Future research aims to embed common‑sense knowledge graphs into the model’s weights, but until then: - Treat agents as powerful autocomplete tools. - Never assume they understand context beyond the supplied prompt. - Keep the human loop tight, especially for actions that affect the real world.
Written by Nur Azhar on 2026‑06‑24. This post reflects personal observations and is not a formal scientific paper.