Curating what enters the context window each turn - the discipline that keeps agents reliable as windows grow. From lean CLAUDE.md files to compaction, subagents, and the agentic loop itself.
Claude Code runs a gather→act→verify loop: startup load fills context before you type, then file reads and tool output pile up as the loop runs - compaction, /clear, and subagents are the official levers for what sticks.
Claude Code runs up to a 1M-token window on flagship models, but a bigger window doesn't fix a cluttered one - every turn replays the whole history and quality drops before the cap.
A bloated CLAUDE.md gets filtered out as noise. Keep it to universal instructions and move everything else behind progressive disclosure.
Managing what a coding agent keeps in context - not window size - cut agent cost over 50% on SWE-bench Verified with no drop in solve rate, per JetBrains Research.
A minimal Claude Code harness fits in ten files - a lean CLAUDE.md skeleton, a verifiable definition-of-done loop, three read-only subagents, a pre-ship security audit, and credential deny rules. One npx command installs the set.
On the one adversarial task its line-36 rule existed to prevent, a 129-instruction CLAUDE.md returned the same one-word reply as a harness with no instructions at all.
One rule, three layers, measured: a rule in Claude Code's auto memory held 3/3 runs until the user contradicted it, then failed 0/3 - the same score the identical rule got in CLAUDE.md.