An old idea gets a new engine

The idea of an agent is older than a language model.

In the late 1960s, Shakey the robot could receive a goal, plan several steps, and move through a simple room. In the 1990s, researchers used the word software agent for programs that acted on a person's behalf.

Those systems had planners, rules, or narrow learned parts. They did not have a general language model that could interpret many tasks through ordinary words.

In 2022, the ReAct paper showed a language model alternating between written reasoning and actions in an external environment. A model could request a search, receive the result as new context, and choose the next step.

The new part was not the wish for a machine that acts. It was the engine inside the loop. A language model could use one text interface to choose among many tools.

The board from the previous module became a workspace for instructions, documents, tool requests, and results.

An old idea gets a new engine

The idea of an agent is older than a language model.

Those systems had planners, rules, or narrow learned parts. They did not have a general language model that could interpret many tasks through ordinary words.

The new part was not the wish for a machine that acts. It was the engine inside the loop. A language model could use one text interface to choose among many tools.

The board from the previous module became a workspace for instructions, documents, tool requests, and results.