18 lines
1.8 KiB
HTML
18 lines
1.8 KiB
HTML
<head>
|
|
<meta name="title" content="Things AI Agents Should Be Able to Do" />
|
|
<meta name="date" content="2025-07-30T00:00:00Z" />
|
|
</head>
|
|
<article>
|
|
<h1>Things AI Agents Should be Able to Do</h1>
|
|
<h2>Write Their Own Helper Scripts</h2>
|
|
<p>
|
|
If I ask my agent to play chess with me (or a game that doesn't exist), it should write a basic chess engine, unmprompted, in a sandboxed environment, which it can then use as a tool the next time it plays chess. Nobody has implemented this yet because it requires a lot of agency from the agent, and would have a bad error rate. I think it can be done effectively by limiting the agent to a single tech stack and a railroaded workflow for creating its own tools. Like a human software engineer, an AI agent should seek to automate tedious tasks.
|
|
</p>
|
|
<h2>Forget</h2>
|
|
<p>
|
|
Humans still have better memories than LLMs, because we forget. Agents can be fitted with a RAG to quickly memorize trivia, but if you have your agent remembering everything it does all day, the database gets cluttered with information that doesn't matter, and recall brings up a lot of garbage. This effect would be exacerbated if you fed an agent a firehose of information from your camera-glasses, which is exactly how much information your own monkey brain is intaking.
|
|
</br>
|
|
Human memory should be used as a model: when information is repeated, we remember it more strongly, and if it's not, we forget. Additionally, humans spend time "alone with our thoughts" reflecting on our memories and creating new super-memories that are much denser with useful information (For example, you might remember "My friend James likes strawberry ice cream," a reflection that allows you to throw away 200 instances of memories of him ordering strawberry ice cream).
|
|
</p>
|
|
</article>
|