Where Codex fails — and why that is normal

Typical agent failures, the cost of weak context, and tasks I no longer delegate.

I use it for larger tasks that need autonomous work across a project, not just autocomplete.

What works in practice

• Do not trust confident wording without checking the code.
• Keep ambiguous product decisions human-owned.
• Stop when the agent starts patching its own patches.
• Restart context when the original hypothesis changes.

Where the boundary is

The main limitation is not the model, but the quality of context and the ability to verify the outcome.