fixes

Author: ollmer

experiments/run_miniwob.py

@@ -1,6 +1,8 @@
 import logging
 import os
 
+from dotenv import load_dotenv
+
 from agentlab.agents.tapeagent.agent import TapeAgentArgs, load_config
 from agentlab.backends.browser.mcp_playwright import MCPPlaywright
 from agentlab.benchmarks.miniwob import MiniWobBenchmark
@@ -9,6 +11,7 @@
 fmt = "%(asctime)s - %(levelname)s - %(name)s:%(lineno)d - %(funcName)s() - %(message)s"
 logging.basicConfig(level=logging.INFO, force=True, format=fmt, handlers=[logging.StreamHandler()])
 logger = logging.getLogger(__name__)
+load_dotenv()
 
 if __name__ == "__main__":
     config = load_config("miniwob")
@@ -20,7 +23,7 @@
         logging_level_stdout=logging.INFO,
     )
     if os.environ.get("AGENTLAB_DEBUG"):
-        study.exp_args_list = study.exp_args_list[:3]
+        study.exp_args_list = study.exp_args_list[:1]
         study.run(n_jobs=1, n_relaunch=1, parallel_backend="sequential")
     else:
         study.run(n_jobs=config.n_jobs, n_relaunch=1, parallel_backend=config.parallel_backend)

experiments/test_mcp.py

@@ -1,6 +1,3 @@
-from tapeagents.environment import FunctionCall
-from tapeagents.mcp import ToolCallAction
-
 from agentlab.backends.browser.mcp_playwright import MCPPlaywright
 from agentlab.benchmarks.miniwob.task import get_miniwob_tasks
 
@@ -11,6 +8,9 @@ def main():
     setup_js = task.get_setup_js()
 
     backend = MCPPlaywright()
+    backend.initialize()
+    print(backend.actions())
+
     print("="*100)
     # 1. goto task url
     print("URL: ", task.url)

src/agentlab/backends/browser/env.py

@@ -29,6 +29,7 @@ def __init__(self, task_name: str, task: AbstractWebTask, backend: BrowserBacken
         self.task = task
         self.seed = seed
         self._turns = 0
+        self.max_turns = task.max_turns
         self.backend = backend
         self.backend.initialize()
 
@@ -45,7 +46,7 @@ def reset(self, seed: int):
     def step(self, action: ToolCallAction | str) -> tuple[Observation, float, bool, bool, dict]:
         if isinstance(action, str):
             action = ToolsActionSet.parse_action(action)
-        logger.info(f"BrowserEnv.step() called with action {action.function.name}")
+        logger.info(f"BrowserEnv.step() called with action {action}")
 
         action_exec_start = time.time()
         finished = isinstance(action, StopStep)
@@ -61,7 +62,7 @@ def step(self, action: ToolCallAction | str) -> tuple[Observation, float, bool,
         if self.task.validate_per_step or finished or truncated:
             reward = self.calculate_reward(action, observation)
         else:
-            reward = None
+            reward = 0.0
 
         env_info = {
             "step_metadata": observation.metadata,

src/agentlab/benchmarks/miniwob/task.py

@@ -16,6 +16,7 @@ class MiniWobTask(AbstractWebTask):
     url: str = None
     remove_human_display: bool = True
     episode_max_time: int = 1000000
+    max_turns: int = 10
     actions_whitelist: ClassVar[list[str]] = [
         "browser_press_key",
         "browser_type",

src/agentlab/benchmarks/web_task.py

@@ -9,6 +9,7 @@ class AbstractWebTask(BaseModel):
     url: str
     validate_per_step: bool = False
     actions_whitelist: ClassVar[list[str]] = []
+    max_turns: int = 100
 
     @classmethod
     def filter_actions(cls, actions: list[ToolSpec]) -> list[str]: