Declarative Pre-Invocation Context Assembly for LLM Agents: A JMESPath-Based Architecture for Deterministic Enterprise Context Fetching

Large language model (LLM) agents deployed in enterprise environments require access to live, structured external data at inference time. Existing approaches i.e retrieval-augmented generation (RAG) and LLM-driven tool calling introduce non-determinism, multi-round latency, and developer-opacity into the context assembly process. This paper presents Pre-LLM Declarative Context Assembly (PLDCA), an architectural pattern in which an agent's definition contains explicit fetch specifications that are evaluated deterministically before any LLM invocation. Each fetch specification declares a data source, an optional authentication method, and a JMESPath expression that transforms and reduces the API response to precisely the tokens the agent requires. At runtime, all fetch specifications are resolved in parallel, their outputs are injected into the prompt template, and a single LLM call is made with a fully pre-assembled context. Compared to tool calling, this eliminates all pre-retrieval LLM round trips; compared to RAG, it replaces probabilistic vector retrieval with developer-authored, auditable data access paths. A worked example on the GitHub Status API demonstrates 92.6% token reduction from a real enterprise API response