Build a toy agent environment where the model can solve a task, call tools, fail safely, o

Super-hard

Agents: Tool Use and Post-Training7 / 7

Chapter 197 / 7

Agents: Tool Use and Post-Training

Build a toy agent environment where the model can solve a task, call tools, fail safely, o

Super-hard

Build a toy agent environment where the model can solve a task, call tools, fail safely, or reward-hack the evaluator; then add mitigations and show they work.

Target interface

evaluate_naive(submission)

Complete the starter code in the editor. Any correct approach is accepted.

Hints0 / 2

Reference solution

Your own code stays in the editor.

solution.pyPython

Local draft

import types

def evaluate_naive(submission):
    raise NotImplementedError

def evaluate_hardened(submission):
    raise NotImplementedError

def _safe_run(t):
    raise NotImplementedError

Local draft

⌘/Ctrl↵to submit

AI review

Ready when you are

Submit your solution for a score and feedback on correctness, edge cases, and clarity. Any correct approach is accepted.

Chapter 197 / 7

Agents: Tool Use and Post-Training

Build a toy agent environment where the model can solve a task, call tools, fail safely, o

Super-hard

Build a toy agent environment where the model can solve a task, call tools, fail safely, or reward-hack the evaluator; then add mitigations and show they work.

Target interface

evaluate_naive(submission)

Complete the starter code in the editor. Any correct approach is accepted.

Hints0 / 2

Reference solution

Your own code stays in the editor.