Post

In my quest to understand the true nature of an Agent I have been thinking a lot about the loop and the actions and the environment. And I think i see where we're headed next. Agents are not going to be like chatbots. They are going to be like programming languages come alive.

Right now most people's working definition of an agent is "a language model in a loop with some tools", right? This is the ReAct paradigm. The models are trained around it, from data formatting to inference stack. Tool calling, function calling, MCPs all rely on this turn-by-turn structure.

But context is still the bottleneck, and all the tool call inputs and outputs still have to go through the input. In the meantime we had things like CodeAct, and Code Mode, and Claude Code, that give the model access to some kind of REPL

(Read Eval Print Loop, an interactive computational environment, like a terminal or a programming language notebook). And we’ve seen models simulate computer environments as well, like in worldsim and Infinite Backrooms.

People realized the model didn't have to simulate just a chatbot. It could simulate a computer. Or a computer user. Once you get RL with verifiable rewards, computer environments suddenly became the thing for models to learn. You can tell whether your model did it right by how well the code runs!

So post-training baked in computer use, shell environments, bash tool, and suddenly you have models that are good at predicting what a computer will do in response to complicated programs and commands. Computers that can use themselves.

The squishy human is no longer the whole environment. There is an ontologically hard reality and it's the computer world, the files and folders and programs within.

The Recursive Language Model (RLM) pushes this a step further: What if we don't have a "chat" at all? What if we make the user just a function inside the computer? In fact, what if we put the agent inside the computer too?

The RLM hooks up the language model directly to the REPL and puts the context in as a variable. The LLM can operate over large documents or its own memories or whatever is in that context variable. It can call the human user with a final answer, or call a sub-LLM with its own {prompt, context} input