Call agents & RL

RL environments & decision engines

An RL rollout needs many identical environments, each one isolated, each one cheap to copy. A Sandbox is that environment: start one, run the episode to a decision point, snapshot its memory, restore as many copies as you have actions to try, and score the trajectories in parallel with .map().

CPU sandboxes for environments, gVisor-isolated with network rules; a GPU function for the policy or the scorer.

Sandbox docs RL Arena

How it works

1Sandbox(...).create() starts an environment; process.run_code drives it step by step and returns what the agent observed.
2At a branching point, snapshot_memory() freezes the whole process; Sandbox.create_from_memory_snapshot(id) brings back an identical copy in seconds, one per candidate action.
3Each copy continues its own episode; the trajectories go to a scoring function, fanned out with .map() so hundreds are judged at once.
4Scores feed the decision engine or the trainer; the sandboxes are terminated, and the next rollout starts from the same snapshot.

The code

rollout.py

from levers9 import Sandbox, job

env = Sandbox(cpu=1, memory="1Gi", block_network=True, keep_warm_seconds=1800).create()
env.process.run_code(open("play_to_decision_point.py").read())
snap = env.snapshot_memory()                       # freeze the episode here

@job(cpu=1, memory="1Gi")
def rollout(action: str) -> dict:
    copy = Sandbox.create_from_memory_snapshot(snap) # an identical environment per action
    out = copy.process.run_code(f"step({action!r}); print(score())")
    copy.terminate()
    return {"action": action, "score": float(out.result)}

results = list(rollout.map(["book_viewing", "ask_budget", "offer_callback"]))

run it

python rollout.py
levers9 container list        # every environment copy while it runs

Where to go next

Sandbox docssnapshots, processes, files, network rules RL Arenathe call-agent loop built on the same idea Containerssandboxes appear here while they run

Sandboxed code execution

Run LLM-generated code in gVisor-isolated containers with CPU, memory and network limits per call.

Teach a call agent with human feedback

Play the seven missions: calls in, hot spots, duels for 10 XP a pick, level up, boss fight, crown the champion, connect it to real customers.

Process ten thousand inputs with .map()

One container per item, sub-second warm starts, scale back to zero when the queue drains.