Time Is the Cheapest Option

Context is the bottleneck, not intelligence. The models are smart enough. What kills them is noise, and the fix is almost always the same operation: compression. Strip what's irrelevant, pass forward what's dense.

Here's the highest-ratio compression I know of, and no harness ships it: the last-modified date.

When an agentic harness reads a file, the agent gets a name and contents. It has no idea whether that file was touched two hours ago or two years ago. But that one fact compresses an enormous amount of judgment into almost nothing. A handoff doc modified 14 months ago probably describes a world that no longer exists. A config edited 20 minutes ago probably explains the bug you're chasing. Humans do this constantly. We sort by modified date before we read a single line. Agents don't, because we never gave them the option.

The token math is absurd. A relative age like (14mo) costs about two tokens. What it buys is the ability to not read things, and skipping a file entirely is infinite compression. It also works exactly where git fails: untracked files, dirty working trees, commit messages that say "fix".

Format matters too. Epoch seconds are byte-compact but force the model into arithmetic it's bad at. ISO dates work. Bucketed relative ages are better, because they're themselves a lossy compression of the timestamp, and everything they throw away is noise. Nobody reasons about staleness at second precision.

The pattern underneath is the same one underneath everything in agent engineering: don't make the model infer what the environment already knows. mtime is a free stat call, a pre-chewed relevance signal, a two-token summary of a file's relationship to now. Compression is comprehension. Time might be the cheapest comprehension we're currently throwing away.