An agent has goals / preferences and makes decisions / takes actions to achieve and to satisfy those goals.
The minimal agent is a controller of future states.
An agent can be something as simple as a thermostat which can control future trajectories, with the goal of keeping the room at a certain temperature. Its decisions affect future trajectories.
In order to be able to do that, it needs to model the world of how the world is going to change as a result of its actions (with different expectation horizons in a branching world, …), rather than just taking measurements. The efficiency of this process is its intelligence.
We end up with a system that seems to have preferences, knowledge, commitments, goal-directedness, …The more an agent is able to model its own actions as a causal factor in the world, the more self-aware (conscious) it becomes.
At this point, you start to edit your own source code, become self-authoring, change the way you interact, …
Transclude of michael-levin#^1ebf7d
Any agent that evolves under some sort of energy and time resource constraint is going to have to coarse grain and not be a reductionist. They have to tell stories about agents that do things as a means of compression, as you cannot track everyhting.
Transclude of company#^b93144
Preferences = desiring to choose the trajectory which minimizes deviations. Decisions = actions.
In the context of machine learning models, “agency” typically refers to the ability of a system to independently and proactively make decisions or take actions.
A model or system with agency can autonomously interpret its environment, make decisions based on that interpretation, and then carry out actions.
It doesn’t merely mean the ability to perform tasks as instructed, but rather it suggests an ability to independently determine which tasks should be performed in order to achieve particular goals. This concept is a key part in reinforcement learning, where an agent learns to make optimal decisions by interacting with an environment. - GPT
papers on agents
VOYAGER - An Open-Ended Embodied Agent with Large Language Models