The U.S. government publishes high-level intelligence budget aggregates — the National Intelligence Program (NIP) and Military Intelligence Program (MIP) toplines — but does not provide a public line-item breakdown of where those funds go. This report describes the Intelligence Budget Estimation System (IBES), a multi-agent workflow that infers the classified "black budget" residual by enumerating observable line items in public Department of Defense (DoD) budget exhibits, classifying them for intelligence relevance, and reconciling them against the disclosed totals. The system is built on LangGraph for workflow orchestration, Pydantic AI for typed agent schemas, and locally hosted large language models served by Ollama, yielding a stack that keeps technology dependencies shallow while making each analytical step explicit and auditable. The current implementation integrates DoD R-1 (Research, Development, Test & Evaluation), P-1 (Procurement), and O-1 (Operation & Maintenance) exhibits across fiscal years 2020–2025, applies a four-tier classification scheme combining rule-based heuristics with LLM reasoning, and produces residual estimates ranging from $82B to $105B per year alongside a linear trend model with R² = 0.836. The report describes the architecture, data integration, classification methodology, results, limitations, and a roadmap for incorporating additional public data sources (SEC 10-K filings, ODNI breakdowns, congressional justifications) to improve coverage.