Things to be Aware of While Using AI Coding Agents

A lot of engineers and engineering organizations have been very successful in quickly embracing AI (agentic) coding tools such as Cursor, WindSurf, Cline and Claude Code. However, we need to keep a few things in mind:

Trust but verify: Do not merge any AI generated code without understanding what it does.
Responsibility: Developers and PR reviewers are both responsible for ensuring that AI generated code is functional, robust and reliable.
Unit Tests: Leverage AI to create as many unit tests as possible. Use mocks as applicable, specially for features that involve UI.
Very targeted prompting: Use AI with very targeted prompting. Do not trying to build a full feature using a single high-level prompt. Be aware of the context window. Plan out your implementation, write it down in a markdown file, and then use AI to help implement step-by-step in a methodical manner.
Planning before coding: Related to above point, Claude Code plan mode should be leveraged as much as possible. Similar mechanism can be used on Cursor even though a dedicated plan mode may not exist there (at least it did not few months back; I have since moved from Cursor to Claude Code).
Managing memory: Related to the aforementioned points about context windows and planning, when using Claude Code, leverage the /memory command. A more custom mechanism is to use sessions. See more here. It is pertinent to mention that working across context windows through compression can be expensive, so planning and executing on small steps, and/or using sessions which can be resumed across context windows is recommended.
Transparency with AI usage: When submitting PRs, be transparent about AI usage.
Use AI as an enabler not replacement (yet). While we certainly want to embrace tools that enable higher productivity, at this point in time I would be wary of losing skills I have acquired over several years [see this (for essays but concept applies) and this]. And I say this even though I am not a puritan. So, read through and fully understand any AI generated code before merging.

Further reading