ConceptReAct / Reasoning and Acting
ReAct Loop
An agent architecture pattern in which a language model alternates between reasoning and tool calls. The model outputs a thought and an action and receives an observation, advancing the task in a loop. The reasoning process is traceable, but errors are cumulative.
Timeline (last 90 days)
Nov 1
When tool results are deliberately not written back into the conversation, the model makes the same request at every step at temperature 0, and the loop repeats until it reaches the maximum step limit and only then terminates; the step limit is the only reason it stops.
Unverified50%