AI agents struggle not with insufficient context but with losing type clarity when instructions, memory, evidence, and tool outputs are flattened into raw strings. This article introduces a lightweight, zero‑dependency Python runtime that implements a context type system, assigning explicit types to each context piece (INSTRUCTION, EVIDENCE, MEMORY, TOOL_OUTPUT) and enforcing transformation rules. The system tracks provenance and rejects invalid changes before the model sees them, ensuring tool output cannot masquerade as an instruction. The author presents the full source code, real terminal output, and eight passing tests without any LLM calls. It serves as a correctness and observability layer—akin to a type checker—rather than adding new model capabilities. The piece targets developers building agent systems with multi‑source prompts, offering a minimalist approach to prevent subtle bugs that masquerade as model failures. It clarifies what the solution does and does not address, directing readers to the GitHub repository for hands‑on exploration.

Key takeaways include the importance of preserving context types, the practical implementation of a type system for agent pipelines, and the value of explicit provenance tracking in debugging complex agent behaviors.

For those constructing multi‑source RAG pipelines, managing tool‑calling agents, or persisting state across turns, this approach provides a clear, testable method to isolate and resolve type‑confusion bugs, ensuring agents behave predictably and transparently.