Comments are the highest-signal channel in an agentic coding workflow not because they’re clever, but because they are unavoidable. Docs are optional, Tool configs are conditional, but a comment sitting directly above the method an agent is about to edit is almost always read. That reliability is the whole game: it’s a direct, in-context instruction that fires at exactly the moment a decision is being made.
What most people miss is how much memory you get for free. A single sentence about intent, constraints, or past landmines becomes durable context that survives across ai providers, agents, sessions, models, and tooling changes with no vector database required. You don’t need special infrastructure to remind an agent why a seemingly clean refactor will break a hidden dependency. You just need a one line comment in the right place.
The discipline in doing this right is keeping comments rare and surgical. A handful of targeted notes such as “keep this in sync with X,” “avoid Y, kills perf at scale,” “Z exists because of legacy clients on v2” etc can prevent hours of backtracking. Overdo it and the signal-to-noise collapses; agents start skimming the same way humans do. But used sparingly, comments function as micro-guides that compound across every agent run touching that file. They’re already embedded where the work happens. That’s the only place they need to be.
