What is an agent?
“Agent” is the most overloaded word in AI right now. It’s marketing copy,
it’s a research term with fifty years of baggage, and it’s the name of the
thing you can build in an afternoon with an LLM and a while loop. Before
this site can teach anything, we need a definition that can bear weight.
The textbook one — an agent is anything that perceives its environment through sensors and acts upon it through actuators — is true but toothless. It admits thermostats. The interesting question is what separates the thermostat from the systems this site is about.
Rather than hand you the answer, let’s do this the Lyceum way:
the definition, in prose
If you steered the dialogue to its end, you arrived here:
An agent is a system that pursues goals over time by taking actions, observing consequences, and adapting.
Each noun in that sentence is a design surface:
- Goals — who sets them, how they’re represented, what happens when they conflict.
- Actions — the tool surface: what the agent is permitted to touch, and with whose authority.
- Observation — what feedback the agent gets, and how honest it is.
- Adaptation — the loop: plan, act, observe, revise. The loop is what makes a model agentic; the model is just the loop’s engine.
And each noun is also an ethical surface. Goals can be misaligned. Actions can be irreversible. Observations can be manipulated — an agent reading a web page is trusting that page. Adaptation without limits is just drift with good branding.
That symmetry — every architectural property doubling as an ethical one — is the thesis of this entire site. The next post in the foundations thread takes the loop apart piece by piece.