Agent workflows evolve as models start splitting their own tasks

Agent automation has moved through three stages, defined by who designs the workflow. The first, Ralph, has no real structure — it just reruns the same prompt in a loop, starting with a fresh context each time and relying on persistence rather than planning. The second, CodeMachine, has a human write the workflow once, and the agent then follows that human-authored structure step by step. The third stage is RLM (Recursive Language Model), described in a paper by researchers Zhang and Kraska at MIT CSAIL: here the model treats its own context as an environment, breaks the task into pieces, and calls itself recursively to handle each piece.

The overall trend is that task-splitting has moved from nobody doing it, to a human doing it, to the model doing it itself. The author built a free, open-source, zero-dependency JavaScript library called bareagent that implements this last stage as a function called recurse(). recurse() works in four steps: decompose the task, fan out into subtasks handled in parallel, verify the results, and synthesize a final answer. The model is given a tool called spawn_child and decides on its own whether to split a task further.

Each spawned child gets its own fresh context window and only sees its specific subtask, then returns either a {result, verdict} pair or an honest {incomplete} if it couldn't finish. The author is explicit that this implements the RLM pattern conceptually, not the exact Python-REPL mechanism described in the original paper.