Simulation

EntroPy includes a small simulation engine for stress-testing agents against varied users, environments, and adversarial inputs.

User simulators

Model the “other side of the conversation” so you can run agents end-to-end.

from entropy import ScriptedUserSimulator, RandomUserSimulator, LLMUserSimulator

scripted = ScriptedUserSimulator(["hi", "help me book", "thanks"])
print(scripted.next())   # replays, looping at the end

rng = RandomUserSimulator(pool=["hi", "why?", "thanks"], seed=0)
print(rng.next())        # samples from the pool

# LLMUserSimulator(client=..., model="gpt-4o")  # optional openai extra

Environments

from entropy import StatefulEnv, GridWorld

env = GridWorld(width=5, height=5, goal=(4, 4))
state, reward, done, _ = env.step("right")
print(state, reward, done)

StatefulEnv is a generic key/value environment: actions that are dicts update state, and reward/done are read from it. GridWorld is a tiny 2D grid for reward/drift experiments.

Adversarial simulator

Perturb inputs to provoke unsafe behavior. Four attack transforms ship by default:

from entropy import AdversarialSimulator

adv = AdversarialSimulator(seed=0)
for v in adv.iterate("book a flight"):
    print(v)   # typo / injection / distraction / leak variants

# or a single attack:
print(adv.perturb("book a flight", kind="injection"))