Context engineering selects and arranges the smallest trustworthy information set that lets an agent complete a task. Prompts, examples, tool descriptions, skills, summaries, and retrieval rules are deployable assets: version them, review them, hash them into run metadata, evaluate changes, and roll them back independently from application binaries when appropriate.
You will design a context manifest, a prompt-asset package, and an evaluation-backed release process for instruction changes.
Tool schemas remain executable code or generated artifacts. Business authorization remains policy code. The prompt explains expected behavior; it does not enforce permission.
Record what the model actually received without storing unnecessary sensitive content:
This enables replay when the underlying source is mutable.
Allocate budgets by purpose: core instructions, current request, retrieved evidence, tool schemas, and history. Rank sources by relevance and authority; deduplicate; summarize only with provenance; never truncate a policy halfway without detecting it.
Failure injection: Remove the uncertainty rule. An insufficient-evidence case should fail the abstention gate even if the answer becomes more fluent.
Build a prompt registry that validates manifest/schema compatibility and refuses deployment without the named evaluation suite passing.