Agent context needs code indexes, retrieval, and durable state

A larger context window does not remove the need to manage an agent's information. Precise code relationships, relevant documents, and saved working state serve different purposes. Each has its own source of truth and invalidation rules. Mixing them can cost more than extra tokens: an agent may confidently act on an obsolete version of the task.

JetBrains' October 15 account describes Netflix using its code-intelligence engine with agents. The same day, Microsoft introduced Python vector-store APIs and context providers, explicitly retaining experimental APIs and alpha or beta integration packages. Vercel announced general availability of Drives for Sandbox. An earlier Merge discussion of company-wide context connects knowledge to freshness, tools, and access. None establishes interoperability among all four offerings.

An index, a search result, and a workspace are different assets

Code indexes answer structural questions: where a symbol is defined, what calls it, and which dependencies surround it. Semantic retrieval finds related language, but textual similarity does not prove a function relationship. A durable workspace retains files between runs; a file's continued existence does not make it an approved deliverable.

That distinction suggests a concrete design. Use a structure-aware tool for code navigation, filtered retrieval with source references for instructions, and explicit records of inputs, decisions, and unfinished actions for continuation. One model may consume all three, while the application preserves their different meanings.

Failures have different consequences. Missing a function can cause extra exploration. An obsolete policy can change the answer. A confused execution state can repeat an external action. A single context-quality score therefore hides the operational questions that matter. Measure the retrieval or transition being supported, rather than treating every added document as equivalent improvement.

This also changes a vendor comparison. Ask which problem a proposed context layer actually solves. A persistent filesystem does not automatically provide semantic search. A searchable collection does not establish an execution checkpoint. A code index does not contain a customer's latest commercial decision. Paying for one capability while expecting all three creates an integration gap.

Durable information needs invalidation rules

Reusing indexes and files reduces preparation work, while creating responsibility for detecting change. An index must match a repository version. A retrieved policy needs a source, owner, and revision. A saved result should identify the inputs that produced it. Faster access without these relationships merely accelerates the reuse of mistakes.

Conversation history requires the same discipline. An assumption made early in a task does not become fact when written to a memory file. Separate confirmed information, working hypotheses, and human decisions. A resumed agent needs those distinctions, rather than a fluent summary that flattens their evidential status.

Access can change faster than an index is rebuilt. Enforce organization and user boundaries when retrieving data, as well as when ingesting it. Durable working state also needs tenant and branch separation. Persistence is a storage property; it does not supply an authorization model by itself.

An implementation should define what happens when a source disappears. Retaining the derived result indefinitely may conflict with the new access state. Deleting it blindly may lose an audit trail. Choose the intended behavior explicitly, including who can inspect provenance after access is withdrawn.

Hypothetical example: changing a pricing rule

Consider a team modifying a pricing calculation in an existing application. This is a design example. The agent uses code intelligence to find calculation paths, retrieval to obtain approved business rules, and durable storage to resume after interruption. Its task record identifies the code revision, rule document, and decisions still awaiting approval.

The customer changes rounding requirements overnight. On resumption, the system should discover the new rule and mark the old calculation for review. Repeating yesterday's test and declaring success would preserve execution while losing the business requirement. Saved files are useful because their origin and remaining validity can be checked.

For a pilot, compare ordinary file search with structural tooling on your own tasks, measuring cold starts separately. Then introduce an obsolete instruction and an interrupted run. A publisher's efficiency result cannot establish the benefit for another language, project size, or tool budget.

Our guide to background workers and retries explains the continuation problem at application level. The Python API data-processing case supplies an integration context. You can discuss the knowledge and state your agent needs before enlarging its context window.