Long-running agents make context an operating expense

Cursor’s September 23, 2026 efficiency report treated context management as an engineering cost problem. Among its changes, the company described more selective use of subagents and reduced repeated prompt overhead. A separate worker can start with a fresh context, but creating that worker also introduces coordination. This makes context a resource to allocate across an entire task, rather than a model specification to maximize in isolation. Cursor token efficiency engineering.

An autonomous workflow often accumulates documents, tool outputs, and discussion. Keeping all of it in every subsequent prompt feels cautious, because nothing is deliberately discarded. Yet a large context can become an expensive substitute for deciding what matters. The economics of memory should be assessed together with the quality of the resulting decisions.

Lost in the Middle, submitted July 6, 2023, found that information position affected performance in the long-context tasks it evaluated. It gives a historical reason to test whether important evidence is actually used. Context capacity alone is not a direct measurement of reliable retrieval. Lost in the Middle paper.

Carry the evidence, not the clutter

Our recommendation is to maintain a concise task state with links back to original evidence. Preserve unresolved questions, confirmed decisions, and the current plan. Keep bulky material available for targeted retrieval. This approach needs testing because a poor summary can remove the very condition that matters; compression should not silently turn tentative information into a settled fact.

FrugalGPT's May 9, 2023, paper discussed prompt adaptation alongside other approaches to managing model-query cost. Its broader relevance is that cost can be influenced by the structure of the request, not only by choosing a cheaper model. Its experimental savings are not current price forecasts. FrugalGPT paper.

Measure useful recall

We would compare full-context and structured-memory approaches on actual completion quality, total inference cost, and recovery from interruptions. Include questions whose answers appear in early tool results, because those test whether the memory policy preserves important details. Efficient autonomy is not the smallest prompt at any cost. It is a workflow that repeatedly brings the right evidence to the next decision.

Anthropic’s design keeps references instead of everything

Anthropic’s September 29, 2025 context-engineering article described using lightweight references—such as file paths and stored queries—to retrieve detail when needed. It also discussed compaction and structured notes for work that exceeds a single context window. These techniques preserve continuity by deciding what to retain and what to fetch again. Their value depends on keeping the information that later decisions actually require. Anthropic context engineering.

Code can keep intermediate data outside the prompt

In its November 4, 2025 MCP article, Anthropic showed how generated code could filter a large spreadsheet before returning a small subset to the model. The same approach could transfer a document between services without reproducing the entire text through successive model calls. The example demonstrates a distinction between data needed for execution and data needed for reasoning. It does not imply that every intermediate result should be hidden from inspection. Anthropic code execution with MCP.

Compression changes the evidence available

The Lost in the Middle research cited below showed that access to a long input does not ensure equally reliable use of every part of it. Modern products have changed substantially since that study, so its historical results should not be treated as current model scores. The enduring issue is the relationship between retained information and the decision that depends on it.

An agent summarizing a repository investigation can retain the conclusion while losing the exception that qualified it. An agent clearing old tool output can remove an error that becomes relevant later. Conversely, retaining everything can bury the instruction that matters. Context management is therefore not simply deletion for economy. It is a decision about what evidence remains available to the next step.

The cost of remembering includes the cost of forgetting

The obvious expense is repeated inference over a growing history. Less obvious is the expense of reconstructing work after an over-aggressive summary. A worker may reopen files, repeat searches, or make a wrong edit because an earlier limitation disappeared. The cheapest prompt at one moment may produce a more expensive task overall.

This explains why fresh subagents are neither a universal solution nor a universal waste. A separate context can isolate a focused investigation and return a concise result. It can also duplicate work or omit the assumptions needed by the parent agent. The trade-off depends on how independently the subtask can be understood and how faithfully its result can be handed back.

For customers, the useful measure is continuity of correct progress. Does the agent resume with the right objective, know what has already been verified, and recognize what remains uncertain? Those properties connect context engineering to the economics of autonomy. Cursor’s recent optimizations and Anthropic’s earlier architecture examples show that providers are actively changing this layer. A larger context window may help, but the product’s decisions about retaining, retrieving, and transferring information remain central to whether long-running work is both affordable and dependable.

The research behind the release

Lost in the Middle, published in July 2023, examined multi-document question answering and key-value retrieval, finding that performance could depend strongly on where relevant information appeared. The work concerned the models and setups tested then, not an immutable limitation with a fixed size today. Its relevance is that nominal capacity and effective use are different measurements. An agent may technically receive a complete record while failing to act on a crucial exception. Context engineering attempts to manage that difference, and a credible evaluation must test whether the retained evidence still supports the decisions made later in the task. Original research paper.