Pass the 10-Second Stuck Test: Proof Your AI Agent Is Truly Autonomous

The Trick
Amjad Masad (@amasad) calls it the 10-Second Stuck Test: if your coding agent freezes for more than 10 s after hitting an error, it's not autonomous—it's an autocomplete toy. (source)

1-Minute How-To

Give the agent a repo with a failing unit test.
Run the full test suite.
Timer starts the instant the first red test appears.
✅ Pass if it: identifies the bug → edits code → re-runs tests → goes green … all inside 10 s of getting "stuck."
❌ Fail if it loops, stalls, or asks for help.

Why It Works
True autonomy = Observe → Plan → Act → Verify. Anything slower hides a human-time tax and burns API credits.

Quick Fixes
• Shrink context: only give the failing module & tests.
• Mock external APIs to prevent network lag.
• Treat prompts like onboarding docs for a junior dev.

Attribution
Concept by @amasad, CEO of Replit. Data pulled from community scrape, Sep 2025.

How to run this with CodeBrain (≈500 chars)

In your privacy-first Obsidian vault, add stuck_test.md with setup + failing test.
Open Claude-Code CLI and run:

code

cb code "pytest -q"

Rube MCP logs output back to the vault and 500+ tools.
Iterate prompts until the agent beats the 10 s timer.
Commit; Google Drive sync keeps everything private yet shareable.