Agenda Intelligence: Building a Trustworthy Geopolitical Intelligence Layer for Agent Workflows

Core Idea: Agenda Intelligence is an MCP server product designed specifically for agent workflows. It enhances the credibility and auditability of LLM outputs through structured request/memo contracts, geographic routing reasoning, and evidence auditing mechanisms, addressing the gaps of general LLMs in evidence discipline and providing auditable strategic risk assessment capabilities.

General Large Language Models (LLMs) are powerful in information synthesis and text generation, but they have fundamental flaws in professional intelligence analysis: lack of evidence discipline, inability to verify sources, susceptibility to hallucinations, and inconsistent output formats. These issues are particularly dangerous in high-risk scenarios like geopolitical risk assessment, potentially leading to major decision-making errors. Agenda Intelligence was created to address these problems.

Agenda Intelligence adopts the "Product Shell" architecture pattern, with the core repository serving as an integration point connecting multiple specialized components:

Layer	Repository	Role
Product Shell	agenda-intelligence-md	MCP server, request/memo pattern, geographic routing, evidence auditing, scoring
Reasoning Method	global-think-tank-analyst	Strategic risk reasoning contract, loaded as the default method for analyze
Vertical Expert	central-asia-caspian-hybrid-intelligence-skill	In-depth expertise in Central Asia/Caspian/Middle Corridor, routed by geography
Vertical Expert	gulf-middle-east-hybrid-intelligence-skill	In-depth expertise in Iran/GCC/Maritime Chokepoints, routed by geography

This layered architecture allows users to automatically route to the corresponding domain experts based on the geographic region of the analysis topic, while maintaining a unified interface contract.

1. MCP Product Shell: The analyze tool accepts structured requests conforming to agenda-request.schema.json, routes them to professionals by geography, and returns memos conforming to agenda-memo.schema.json, ensuring input/output predictability.
2. Markdown Protocol: Agenda-Intelligence.md defines the structured reasoning workflow for agents, standardizes analysis steps and output formats, and is both human-readable and machine-parsable.
3. JSON Schema System: Maintains a complete set of JSON Schemas, including agenda-brief.schema.json (briefing), evidence-pack.schema.json (evidence pack), evidence-audit.schema.json (evidence audit), etc.
4. CLI Toolset: Provides over 30 command-line tools, such as doctor (status report), validate-brief (schema validation), score (heuristic scoring), audit-claims (claim auditing), etc.
5. MCP Server: The stdio MCP server exposes 16 tools, covering the validation layer and product layer.

The core value of Agenda Intelligence lies in emphasizing evidence discipline and implementing strict source policies: each claim must have clear source labels (Axis A/B), and source requirements are defined for 12 categories. Key principles:

1. Not a factual verifier: Only checks structure and evidence labels, does not verify the objective truth of claims.
2. Not an autonomous news agent: Does not actively retrieve real-time information, only processes user-provided inputs.
3. Not a source reputation scorer: Does not evaluate source credibility, only verifies whether citations are correct.

Evaluation: The project includes 20 source-supported test cases, which can be reproduced using the agenda-intelligence bench command:

Metric	Value
Number of Cases	20
Average Score	87.6/100
Min/Max	84/91
Schema Validation Pass Rate	100%
With Evidence Pack	100%
With Claim-level Audit	100%
Average Source Coverage	14.8%
Cases with Source Coverage Gaps	20
Orphan Evidence Citations	0

It should be noted that heuristic scores are uncalibrated and not validated against expert judgments. They assess structure, evidence labels, source coverage diagnosis, and decision readiness, not factual truth.

Use Cases:

• Quick Start: Install via pip install agenda-intelligence-md and use commands like doctor, validate-brief, score, etc.
• MCP Client Integration: Supports clients like Cursor; generate configuration via agenda-intelligence mcp-config --client cursor then call the tools.
• Full Analysis Tracking: Provides end-to-end analyze process examples with reproducible scripts.

Limitations: The system does not provide functions such as factual verification, autonomous information retrieval, source reputation scoring, replacement of analyst judgments, compliance/legal/financial consulting, etc.

Future Directions: Integrate real-time information retrieval (while maintaining evidence discipline), develop source reputation assessment modules, expand geographic coverage, establish a community best practices library, and develop a visual interface to lower the barrier to use.

Agenda Intelligence demonstrates how to leverage LLM capabilities while addressing their inherent limitations through architectural design. It does not make LLMs more "intelligent" but rather makes their outputs more reliable, auditable, and suitable for professional scenarios through structured constraints and validation mechanisms. This "constraints over capabilities" design pattern provides valuable references for AI applications in other fields, especially for enterprise intelligence analysis, risk assessment reports, multi-agent collaboration, compliance-sensitive scenarios, etc.