That purely depends on how you write comments.
Things that are really useful in comments:
- URLS to llms.txt or other authoritative docs so agents see it, read it
- Related entities that may require multiple hops to deduce; you can short circuit the agents read by directly pointing your comment to a related entity that may not show up directly in the call path (e.g. separated by a message queue)
- Reasoning and "why" a decision was made so in a future pass, an agent understands why a condition or rule exists
You can write bad comments, but you can now just prompt/instruct your agents to write good comments and use it as long-term memory.