Rollout
Running an agent through a task once to see what it does — a single sampled attempt from start to finish.
Think of it like
A dress rehearsal: let it play the whole thing through once and watch how it goes.
Example
To gather training data, a team generates thousands of rollouts of an agent solving math problems, then keeps the ones that reached the right answer.
How it actually works
A rollout is one full sampled run of an agent’s policy, producing a trajectory. The word comes from reinforcement learning, where you roll out the policy to collect experience, then learn from which runs succeeded. Because sampling is random, rollouts vary — the same task can produce a brilliant run and a hopeless one — which is why people generate many and score them.
For product teams
Why agents feel inconsistent — each run is one roll of the dice, not a fixed program.
For engineers
A single sampled execution of a policy from initial to terminal state, yielding a trajectory for evaluation or training.
Related
- Trajectory — The record a rollout produces.
- RLHF — The learning method the term comes from.
- Temperature — The randomness that makes rollouts differ.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome