This seems to argue for a single agent loop, rather than spread across multiple agents:
> Even in an agentic loop, it's tempting to organize task execution into specialized subagents. Doesn't it feel so clever to come up with these architectures? "The Widget CEO will delegate work to the Widget Engineer, and then the work will be verified by the Widget Tester, with input from the Widget Product Manager".
> While the idea is smart, the results are… dumb. Context is everything for an LLM - when every layer of abstraction introduces loss of context, the ability to string tools together and self-correct course washes away. Subagents still have their place...
Is this really true, if context is carried across? The same quote later refers to Claude and I thought Claude used many subagents, even with AIs writing prompts for other AIs.
I ask mainly because I have built an agent architecture based on state loops, as in their diagrams. (A coding agent: code, then a code review agent, step by step through todos.) Context is communicated across but the idea is wholly to have dedicated subagents. Now I wonder if this is a bad idea.
In my experience building AI Agents, having more specialized agents works better than one big agent. Simply because LLMs are dumb and make A LOT of mistakes. The longer the system prompt and more instructions you give it, the more likely it will miss some of them.
For example, in one architecture, I used to have a support agent reply directly in the customer's language, but the translations were very poor. Now I have one agent thinking of the answer in English, and then one dedicated agent for translation. The translations are A LOT better now.
The main reason why we advocate for a single-loop is because subagents suffer from the telephone game problem: context is lost when the interface is just an agent-to-agent prompt.
We are still working on subagents for parallelization, but if you can stick to the main loop, everything works much better.
> Even in an agentic loop, it's tempting to organize task execution into specialized subagents. Doesn't it feel so clever to come up with these architectures? "The Widget CEO will delegate work to the Widget Engineer, and then the work will be verified by the Widget Tester, with input from the Widget Product Manager".
> While the idea is smart, the results are… dumb. Context is everything for an LLM - when every layer of abstraction introduces loss of context, the ability to string tools together and self-correct course washes away. Subagents still have their place...
Is this really true, if context is carried across? The same quote later refers to Claude and I thought Claude used many subagents, even with AIs writing prompts for other AIs.
I ask mainly because I have built an agent architecture based on state loops, as in their diagrams. (A coding agent: code, then a code review agent, step by step through todos.) Context is communicated across but the idea is wholly to have dedicated subagents. Now I wonder if this is a bad idea.