Skip to content
ART

Strands Math Agent

An RL-trainable GSM8K math agent using Bedrock AgentCore RL Toolkit. The agent solves grade-school word problems by invoking a calculator tool for arithmetic steps and emitting a final answer after ####.

This is the “hello world” for the toolkit — minimal prompt and single tool. Start here.

Quickstart

Section titled “Quickstart”
Terminal window
cd examples/strands_math_agent
uv venv --python 3.13 && source .venv/bin/activate
uv pip install -e .


# Run the RL-adapted app as a local HTTP server
python rl_app.py

Deploy the same entrypoint to AgentCore Runtime via the Prepare agent for RL → Deploy flow, then feed the resulting runtime ARN into SlimeRunner to train.

What’s in the example

Section titled “What’s in the example”
  • basic_app.py — deployment-ready reference using BedrockModel.
  • rl_app.py — RL-adapted version; reads base_url / model_id from _rollout, returns {"rewards": ...}.
  • reward.py — GSM8KReward: extract the #### answer, exact-match against ground truth.
  • evaluate.py — batch-evaluate a deployed ACR agent via RolloutClient.

For local setup (Bedrock credentials, local vLLM, S3 bucket, sending curl requests), Docker builds, ECR push, IAM setup, and full deploy instructions, see the full README on GitHub.