Strands AppWorld Agent

An RL-trainable AppWorld agent using Bedrock AgentCore RL Toolkit. The agent solves day-to-day tasks by interacting with simulated app APIs (Spotify, Venmo, Gmail, etc.) via a Python REPL.

The agent uses a ReAct code agent pattern: a single execute tool runs Python code inside an AppWorld sandbox. APIs are discovered on-demand through the pre-loaded apis object, so tool context stays small regardless of how many APIs exist.

Quickstart

cd examples/strands_appworld_agent
uv venv --python 3.12 && source .venv/bin/activate
UV_GIT_LFS=1 uv pip install -e .
appworld install && appworld download data

# Start the RL app (needs a local vLLM server running on :4000)
python rl_app.py

Deploy using the commands in the AppWorld README, then evaluate with evaluate.py or train with the slime backend.

What’s in the example

• rl_app.py — the rollout entrypoint; spins up an AppWorldServers + AppWorld context per-session and binds the execute tool as a closure over the task’s world.
• reward.py — AppWorldReward: pass-rate from the task’s test_tracker.
• evaluate.py — async batch evaluation over the AppWorld splits (train / dev / test_normal / test_challenge).
• few_shot_example.py — few-shot demo messages seeded into the agent’s context.
• Dockerfile — container dependencies and AppWorld dataset.
• config.example.toml — Runtime ARN and evaluation settings.

For AppWorld data download and evaluation options, see the full README on GitHub.