Agents Were LLMs All Along.
Thoughts on Future of AI Agents
Here are few of my raw thoughts that I posted on socials. I will get back to this post and update as I do more thinking about this topic. :)
Agents were LLMs all along.
All the agentic capabilities are now getting absorbed into the capabilities of LLMs themselves. In 2023, I remember agents being called the next "killer app", and defined loosely as LLMs with tools, memory, reasoning..
Now more of that is built into the LLMs. What's coming is:
- compound systems (e.g., Groq putting more components such as RAG tgthr with LLM)
- cognitive cores (highly capable personalized small LLMs)
- multi-LLM systems instead of multi-agent systems with roles like planner, executor, analyst agent etc., often just wrapping one model in multiple behaviors. (Switching between models also got cheaper.)
Even products that position themselves as agents (like Manus) are becoming more general and running variety of tools (using E2B as the cloud computer for that).
I think for agentic companies, this shift means they’ll win not by gluing tools to LLMs, but by smartly orchestrating multiple models, designing better UIs and workflows, and offering real customization and control on top of the LLMs/agents/however you call it.