agent

Definition

A system that chooses actions toward a goal, observes the outcome, and decides what to do next. With LLMs this is typically a loop wrapping the model with tool calls, memory, and a termination condition. The hard part is not the loop but error compounding: unlike single-turn use, a wrong action changes the state the agent works in, so mistakes propagate instead of being independently sampled.

Explainers using this term

• Paper Walkthrough: Terminal-Universe — Turning Agent Logs Back Into Reusable Execution EnvironmentsTerminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments
• HarnessDev explained (arXiv:2609.01437) — agent harness benchmarkHarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?
• Paper Walkthrough: EarlyEval — Making Agent Evaluation Cheaper by Stopping EarlyEarlyEval: Cheaper Agent Evaluation via Early Outcome Prediction