Get the latest tech news
Inner Loop Agents
What if an LLM could use tools directly? In this post I discuss a potentially divergent view of agents, where agents are less like systems and more like LLMs specially trained to solve problems with a specific set of tools.
If you’ve gotten a chance to use an agent, like Deep Research or o3, you’ll notice that it’s thought process isn’t just inner dialog, it’s also tool calls like web searches. Web search & bash usage are cool and all, but what would be truly powerful is if one of these inner loop agents were trained to use tools that regular people use. I hope it’ll happen, but it seems unlikely that an LLM can discover the finer points of how to use a tool just from reading the manual, without any training.
Or read this on Hacker News