strands-env 0.5.1

A framework for building agent environments with Strands Agents.

strands-env

A framework for building agent environments for RL training and evaluation with Strands Agents.

Features

An agent environment takes a task and runs the agent to completion over multiple turns, producing a rollout result — the trajectory, reward, and termination reason for that task. With strands-env, you can:

• Define Environments — Subclass Environment, add @tool functions, plug in RewardFunction
• RL Training — Token-level trajectories (TITO) for on-policy training with strands-sglang
• Benchmarking — CLI and Evaluator with checkpointing, resume, and custom metrics

Install

pip install strands-env

For development:

git clone https://github.com/horizon-rl/strands-env.git && cd strands-env
pip install -e ".[dev]"

Quick Start

Define an Environment

Subclass Environment and add tools as @tool-decorated functions:

import subprocess
import sys

from strands import tool
from strands_env.core import Environment

@tool
def run_python(code: str) -> str:
    """Run a Python snippet and return its output."""
    proc = subprocess.run([sys.executable, "-c", code], capture_output=True, text=True, timeout=10)
    return proc.stdout + proc.stderr

class CodingEnv(Environment):
    def get_tools(self):
        return [run_python]

Run It

from strands_env.core import Task, TaskContext

env = CodingEnv(model_factory=factory, reward_fn=reward_fn)
result = await env.rollout(Task(
    message="Write Python to compute the 10th Fibonacci number, then run it.",
    context=TaskContext(ground_truth="55"),
))

result.final_response       # "The 10th Fibonacci number is 55"
result.reward_result        # {"reward": 1.0, "info": ...}
result.termination_reason   # TerminationReason.TASK_COMPLETE

See the examples/ directory for complete, runnable demos.

Run Evaluations

python -m strands_env.eval \
    --benchmark terminal-bench-2 \
    --env examples.eval.terminal_bench.terminal_bench_env \
    --backend sglang \
    --base-url http://localhost:30000 \
    --n-samples-per-prompt 4 \
    --max-concurrency 8

Raise --n-samples-per-prompt for more stable pass@k, and --max-concurrency if you're using a hosted sandbox service.

Tip: For a non-agentic benchmark (no tool use), don't override get_tools() — the base class returns [] by default.

Built-in Environments

Ready-to-use environments under src/strands_env/environments/. Each ships with its own README, system prompt, and requirements.txt.

Environment	Description
calculator	Simple environment with a calculator tool for math reasoning.
harbor	Run Harbor-format tasks in sandboxes. Supports training like SETA and evaluation like Terminal-Bench and SWE-bench.
agentcore_code	Python / shell execution via AWS Bedrock AgentCore Code Interpreter.
web_search	Google search + Jina page scraping with optional LLM summarization, enlightened by OpenSeeker.
mcp_atlas	MCP-Atlas benchmark runner across 36 MCP servers with 500 tasks.
agent_world_model	AgentWorldModel tasks with 1000 synthetic FastAPI + SQLite environments exposed as MCP tools.

Documentation

• Evaluation Guide — CLI reference, hook files, custom evaluators
• RL Training Integration — integration with the slime RL training framework

Development

# Lint
ruff check src/ && ruff format --check src/

# Unit tests
pytest tests/unit/ -v

# Integration tests (requires running SGLang server)
pytest tests/integration/ -v --sglang-base-url=http://localhost:30000

Or if using Claude Code, just use /run-unit-tests and /run-integration-tests slash commands.

License

Apache License 2.0 — see LICENSE.