Quiz: Nano Harness and Agent Internals

Test your understanding of agent loops, tools, and sandboxing.

选项

• A. Call model → return result immediately
• B. Call model → parse Python code → execute with tools → observe → repeat until final_answer()
• C. Model returns JSON → parse → store in database → done
• D. Keep calling the model until it returns the word 'done'

答案解析

• A — That's not an agent loop. A loop requires observation and repetition.
• B（正确答案） — Correct! That's the core loop: model → code → execute → observe → repeat.
• C — That's a one-shot API call, not an agent loop.
• D — The agent needs to execute code, not just call the model repeatedly.

选项

• A. The model outputs English text like 'The answer is 42'
• B. The model outputs Python code that calls tools; agent executes that code in a sandbox
• C. The model outputs JSON which the agent parses
• D. The model outputs commands like 'list_dir' and the agent translates them

答案解析

• A — Nano Harness is code-first. Model outputs Python code, not prose.
• B（正确答案） — Correct! Model → Python code → execution → observation.
• C — Some models do this, but nano_harness expects Python code.
• D — The model directly outputs Python function calls.

选项

• A. The message history grows with every loop iteration (model output + observation)
• B. The message history is cleared every step to save tokens
• C. The message history is stored in a database for later analysis
• D. The message history doesn't matter; only the latest model output matters

答案解析

• A（正确答案） — Correct! Each step adds assistant response + user observation. History provides memory.
• B — False. Full history is kept so the model can learn from prior steps.
• C — The history is maintained in memory during the loop.
• D — False. History provides crucial context for the model to make good decisions.

选项

• A. safe_path() allows reading any file on the system
• B. safe_path() prevents directory traversal by checking that paths stay within WORKSPACE
• C. safe_path() is optional; security isn't really needed
• D. safe_path() encrypts paths for transmission

答案解析

• A — False. safe_path() confines paths to the workspace directory.
• B（正确答案） — Correct! It resolves paths and ensures they're within the workspace.
• C — Safety is critical. Without safe_path(), agents could read /etc/passwd.
• D — No, it just prevents path escape attacks.

选项

• A. exec_cmd() allows any shell command to run
• B. exec_cmd() maintains a whitelist of safe commands; anything else is blocked
• C. exec_cmd() has a timeout to prevent hanging commands
• D. exec_cmd() runs commands with admin privileges

答案解析

• A — False. Only whitelisted commands (ls, cat, pwd, etc.) are allowed.
• B（正确答案） — Correct! Only safe commands like 'ls' work. 'rm' and 'curl' are blocked.
• C — True, but the whitelist is the primary security control.
• D — No. Commands run with normal user privileges and are rate-limited.