token-efficiency

Token efficiency

The scarce resource in a long agent session is rarely compute — it's the context window, and by extension the model's ability to keep the actually load-bearing facts in view. Every unnecessary token read into context is a token that isn't a real requirement, a real file, or a real decision competing for attention later in the same session. Optimize for signal density, not for doing less work.

Read narrowly before reading broadly

• Grep before Read. If you know roughly what you're looking for — a function name, an error string, an import — search for it first and read only the matching region, not the whole file. A 2,000-line file read in full to find one 20-line function wastes the other 1,980 lines of context for the rest of the session.
• Use line-range reads on files you already know are large. Most read tools accept an offset/limit; use it once wc -l or a directory listing tells you a file is big.
• Don't re-read what you just wrote. An edit tool that succeeded already confirms the change landed — reading the file back to verify is usually pure waste unless something downstream depends on the exact resulting bytes.
• Prefer targeted tool output over dumping raw command output. git diff --stat before git diff; git log --oneline -20 before an unbounded log; a filtered grep before cat-ing a whole directory.

Delegate what you don't need to keep

• Fork or subagent for research whose conclusion matters more than its trail. If you need an answer — "what does this module do," "where is X defined," "is there a skill for Y" — but not the search process that produced it, hand it to a subagent (or a fork, when one is available) and consume only its final summary. This is the single biggest lever: a subagent's tool noise never enters the parent's context at all.
• Batch independent lookups into one round of parallel calls rather than serial back-and-forth, so the model isn't re-stating context between each one.
• Push large intermediate artifacts to disk, not to the transcript. Write scratch analysis, generated data, or long intermediate output to a file and reference the file path, rather than pasting the full content into the conversation for a later step to re-read anyway.