Stop the agent if it is still running, preserve the current working state, and inspect the complete diff before reverting anything. Keep changes required for the task, restore unrelated edits from a known checkpoint or version control, then tighten the next run’s file boundaries, permissions, and approval rules. The agent’s final chat message is not a substitute for reviewing what it changed.
Stop the active run without erasing evidence
If the agent is still making changes, interrupt or cancel its current turn. Do not immediately reset, clean, or discard the working tree: it may contain useful task work, unrelated edits, or changes you made before starting the agent.
Preserve the current state first. If practical, record the current diff or create a temporary checkpoint before attempting recovery. Codex CLI documentation recommends steering an active turn, inspecting commands and diffs as they appear, and keeping follow-up work in the same session; it also recommends Git checkpoints before and after a task. Interface details vary by agent. OpenAI’s Codex CLI documentation
Find every change and compare it with the request
Review the complete working-tree diff and the full list of changed files, including untracked files where your tools show them. Compare each change with the requested outcome: would the task still be complete without this edit? If not, identify why it is necessary before deciding to keep it. Check files the agent did not mention in its summary as well as the files it says it changed.
#1 Best Overall
Use a clean baseline or checkpoint as the comparison point when available. If the working tree already had uncommitted work before the run, distinguish that work from the agent’s edits before restoring anything; a broad reset can destroy both.
Codex CLI guidance recommends inspecting diffs and commands rather than relying only on the agent’s account of its work. The Codex CLI documentation
Keep task edits and revert scope creep
Separate changes needed for the requested result from changes that are merely adjacent, convenient, or unexplained. Keep the necessary edits; restore unrelated ones from the checkpoint or with version control. Review each restoration against the baseline, especially if you had your own uncommitted changes before the agent started. If you cannot tell whether an edit is safe to remove, preserve it and investigate rather than deleting it blindly.
For future tasks, create a checkpoint before the run and another after reviewing the result. That gives you a known recovery point without treating every change in the working tree as disposable. Codex CLI documents Git checkpoints as a way to revert changes. OpenAI’s Codex CLI documentation
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsWhy a coding agent may edit unrelated files
A coding agent can inspect files, write changes, run commands, and carry out multi-step tasks. A broad request, wide filesystem access, or permissive command approvals can give it room to take actions beyond what you intended. That does not establish why a particular edit happened; the diff and, where available, the command and approval history are better evidence.
Defaults are product-specific. GitHub documents that Copilot CLI filesystem access is scoped by default to the directory where it started, while prompts depend on the active permission mode. Optional computer-use capability can interact with desktop applications beyond that directory boundary. Do not assume another agent has the same limits, or that a directory boundary covers every tool it can use. GitHub’s Copilot Agents documentation
Rank #3
Make the next request and its boundaries explicit
Describe the intended outcome and the permitted files, directories, or subsystems. Name important exclusions, and tell the agent what to do if it believes work outside those limits is necessary: stop and ask before proceeding. Where supported, ask for a plan or confirmation before edits begin.
- Outcome: State the behavior or result you want, not just a vague request to improve something.
- Allowed scope: Identify the files, folders, or subsystem the agent may change.
- Exclusions: Name nearby files, generated assets, configuration, or other areas that should remain untouched.
- Escalation: Require the agent to explain and seek approval before crossing the boundary.
- Review: Ask for a summary of changed files, but verify it against the full diff yourself.
Limit permissions and require approval where it matters
Use the narrowest filesystem boundary, working directory, and tool access that still lets the agent finish the task. Require approval for ambiguous or consequential actions, and avoid broad automatic approvals when they are unnecessary. Check whether approval is requested once or applies across a session: an approval that persists can authorize later commands you did not specifically consider.
GitHub’s Copilot CLI documentation describes one-time and session-level approvals and warns that approving a command such as rm for a session could allow a later rm -rf without another prompt. It recommends sandboxed execution to mitigate risks from automatic approvals. Permission behavior depends on the active mode and configuration. GitHub’s Copilot CLI documentation
For organizations, permission controls should cover what an agent can access, when human approval is needed, which systems it can interact with, and what activity is logged. OpenAI’s safety guidance describes controls for access, approvals, network, identity, rules, and telemetry. Running Codex safely at OpenAI
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Review the result before accepting it
Before merging, committing, or otherwise treating the task as done, inspect the entire diff against the request. Run checks appropriate to the project, and confirm that the final result does not include unrelated edits or unexpected side effects. A concise summary from the agent can help you navigate the changes, but it cannot prove that the full set of changes is in scope.
For teams building a custom agent workflow
Enforce scope where a side effect occurs: at the tool that writes a file, runs a command, or otherwise changes a system. A prompt or final-output check alone may not inspect every tool action, particularly across a multi-agent workflow. Validate a proposed action against the written scope, and pause ambiguous or high-risk actions for human approval.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →OpenAI Agents SDK guidance notes that input guardrails run only for the first agent, output guardrails only for the final agent, and tool guardrails only on tools to which they are attached. It recommends placing validation next to the tool that causes the side effect. OpenAI’s guardrails and human review guidance
For auditability, record relevant prompts, approval decisions, tool executions, and outcomes where your setup supports it. OpenAI describes telemetry examples that include prompts, tool approval decisions, tool execution results, MCP usage, and network allow/deny events. Running Codex safely at OpenAI
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
