-
Compacting context as a session nears its limit
As a long agent session nears its context-window limit, summarize older turns (and/or clear stale tool-call outputs) into a condensed form and continue, rather than truncating blindly or letting the session fail outright.
Why this verdict?
Anthropic lists compaction alongside just-in-time retrieval and note-taking as a standard technique for running agents past a single context window; clearing stale tool results is called out as especially low-risk since raw outputs are usually only needed briefly after the call. Caveat: recent research on a 'compaction cliff' finds naive summarization degrades sharply past a point, so what's kept vs. dropped should be deliberate, not automatic.
-
Just-in-time context retrieval over upfront loading
Instead of pre-loading all potentially-relevant data into context, the agent keeps lightweight references (file paths, IDs, query params) and uses tools to fetch only what it needs at the moment it needs it -- e.g. Claude Code using grep/head/tail against a codebase instead of loading whole files.
Why this verdict?
Anthropic's 2026 context-engineering guidance names this as a core technique for long-horizon agents, alongside compaction and note-taking: preloading everything burns context budget on data that's never used and crowds out tokens the model needs for actual reasoning, mirroring how people use indexes/filesystems instead of memorizing whole corpora.
-
Small, scoped skills loaded on demand
Splitting instructions into small, task-scoped files (skills) that are only loaded into context when relevant, instead of one static mega-file read every turn.
Why this verdict?
Keeps the model's working context focused on the task at hand, lets instructions be authored and updated independently, and scales to far more accumulated knowledge than a single file can hold.
-
One giant CLAUDE.md read every turn
Putting all project context, conventions, and instructions into a single large CLAUDE.md (or system prompt) that is read in full on every turn.
Superseded by
progressive-scoped-skillsWhy this verdict?
A single sprawling context file crowds out the model's working context, goes stale as the project changes, and can't be selectively loaded -- Anthropic's own engineering guidance now recommends small, scoped files/skills loaded progressively instead.