Write, Select, Compress, Isolate

Context management reduces to four moves: write state out, select what to load, compress what is loaded, and isolate work across sub-agents, sandboxes or partitioned state.

Evidence: moderate

In June 2025 Andrej Karpathy publicly backed "context engineering" over "prompt engineering", describing it as the art and science of filling the window with the right information for the next step. The phrase names the problem (context engineering) but not the toolkit. Once context rot is accepted, a team needs a short list of moves it can make.

Lance Martin's "Context Engineering for Agents" supplies four:

  • Write state out of the window, into scratchpads and long-term memory.
  • Select what to bring in: retrieve memories, tools and knowledge, including retrieval over tool descriptions.
  • Compress what is loaded, by summarising or trimming; Claude Code, for example, auto-compacts at about 95% of the window.
  • Isolate work across sub-agents, sandboxes or partitioned state, so each context holds only what its task needs.

Martin draws on Breunig's four context failure modes, on Cognition's multi-agent arguments and on Anthropic's multi-agent work. The list is a practitioner synthesis rather than a measured result, but it organises most agent context practice. Writing out is what a session log outside the context window and externalised progress artefacts do, and the long-term side of it is memory management. Selecting is progressive disclosure and retrieval. Compressing is compaction, with its trade-offs in compaction vs context reset. Isolating is what sub-agents are best at, provided writes stay single-threaded. Google's alternative framing, building each call's context from durable state, is in context as a compiled view.

The four moves also work as a design checklist for any long-running agent: write task state to a file or log rather than keep it in conversation; select log excerpts and reference sections on demand; compress only the working context, never the raw evidence; and isolate read-heavy investigation (for example one sub-agent per subsystem) from the single actor allowed to change anything.

Sources

  • Lance Martin, LangChain, 23 Jun 2025 — https://rlancemartin.github.io/2025/06/23/context_engineering/
  • Andrej Karpathy, June 2025 — https://x.com/karpathy/status/1937902205765607626