Guides

Goal

open the Amazon shopping website

TL;DR

The agent failed to open the Amazon website, and instead initiated a product search for "CCTV cameras", got stuck in repeated search and scroll loops, and continued interacting after incorrectly marking the task as complete.

What Happened

• Step 8-9: Instead of opening the app/website, the agent immediately clicks the search bar and types "CCTV cameras under 5000 including mini cameras".
• Steps 13-15: The agent repeats the exact same search action, indicating it's stuck in a loop.
• Steps 23-28: The agent gets stuck in a scroll loop, executing the same drag action six consecutive times.
• Step 29: The agent incorrectly declares the task finished with a summary of the search results, despite this not being the user's goal.
• Steps 30-32: The agent continues to click and press back after declaring the task finished, demonstrating an overrun.

Root Cause (Best Guess)

The primary failure is in goal comprehension. The agent appears to have ignored the user's simple instruction and substituted a more complex task, possibly from a prior context. The subsequent loops indicate a failure in state detection, where the agent doesn't recognize its actions are not progressing the task.

Where to Look

1. Review the intent recognition and goal parsing logic to understand why the simple goal was misinterpreted.
2. Investigate the state-diffing mechanism that should prevent action loops when the UI is not changing.
3. Check the logic that should terminate the session after a finished step is emitted.

Repro

1. Task: "open the Amazon shopping website"
2. Observe step 8 — the agent starts searching for a product instead of just opening the site.
3. Pattern: The agent falls into loops (e.g., steps 13-15 and 23-28) and overruns the task after an incorrect finished step (steps 30-32).