Skip to content

README.md

Latest commit

 

History

History
73 lines (51 loc) · 3.34 KB

README.md

File metadata and controls

73 lines (51 loc) · 3.34 KB
Raw

AgentWorldModel Environment

MCP environment for AgentWorldModel tasks — 1000 synthetic FastAPI + SQLite environments exposed as MCP tools via fastapi_mcp.

Setup

Install additional dependencies:

pip install -r requirements.txt

This installs AgentWorldModel (agent_world_model) for scenario generation and verification.

Usage

from strands_env.environments.agent_world_model import AgentWorldModelEnv, AgentWorldModelTask

env = AgentWorldModelEnv(
    model_factory=model_factory,
    envs_path="/path/to/gen_envs.jsonl",
    max_tool_iters=10,
)

task = AgentWorldModelTask(
    message="Add two apples to the cart and report the total.",
    scenario="your_scenario",
    task_idx=0,
    verify_code=verify_code,  # from gen_verifier.pure_code.jsonl
    initial_db_path="/path/to/initial.db",
)
result = await env.rollout(task)  # reset (clone DB + start server) -> episode -> reward -> cleanup

AgentWorldModelReward is built in — it is tied to the env's working DB, so there is no reward_fn parameter.

Configuration

Field Default Meaning
envs_path required Path to gen_envs.jsonl (contains scenario, db_path, full_code)
tool_call_timeout 60 MCP tool call timeout in seconds

Base knobs (system_prompt, max_tool_iters, ...) come from EnvironmentConfig.

Task Fields

AgentWorldModelTask carries the per-sample payload:

Field Meaning
scenario Scenario name in gen_envs.jsonl (which synthetic world to boot)
task_idx Task index within the scenario (used in logs)
verify_code Python source defining verify_task_completion(...)
initial_db_path Pristine SQLite snapshot; the episode's working DB is cloned from it

Tools

Dynamic — discovered per episode from the scenario's FastAPI server via MCP (fastapi_mcp); each endpoint of the synthetic world becomes one tool.

Reward

AgentWorldModelReward runs the task's verify_task_completion(initial_db_path, final_db_path, final_answer) via exec(). Each scenario has a unique verification function (from gen_verifier.pure_code.jsonl) that checks:

  • DB state changes — compares the initial snapshot against the episode's working DB (e.g. "was the item added to cart?")
  • Agent's final answer — validates RolloutResult.final_response (e.g. "is the reported total correct?")

Returns 1.0 if result["result"] == "complete", 0.0 otherwise. info["status"] is "success" whenever the verifier ran (reward 0.0 = the agent failed the task) and "error" only for verification-machinery failures — the same contract as the other environments.

Lifecycle

  • reset(task) — Creates a fresh scratch dir (mkdtemp), clones task.initial_db_path into it as the working DB (per-episode isolation by construction — concurrent episodes never share mutable state), generates and starts a FastAPI server subprocess on the clone, waits for TCP readiness, opens an MCP session, discovers tools.
  • rollout(task) — Runs the full episode: reset, the agent loop against the MCP tools, reward, cleanup.
  • cleanup() — Clears tools, closes the MCP session/transport, kills the server process group (SIGTERM, then SIGKILL), removes the scratch dir it created.