Agenda Intel MD adds schema checks for strategic-risk briefs

Agenda Intel MD is an open-source evidence and audit layer for strategic intelligence agents. The project combines a markdown protocol, JSON schemas, a command-line tool, and an MCP-compatible server to help AI systems produce strategic-risk briefs that are structured, evidence-labeled, and easier for analysts to review.

According to the project description, the toolkit is aimed at engineers building policy, sanctions, regulation, geopolitical-risk, market-risk, and strategic-intelligence agents. Its goal is to move agent output away from unsupported summaries and toward briefs that clearly state what changed, why it matters, what is backed by evidence, what remains uncertain, who gains or loses use, which scenarios are plausible, and what to watch next.

The repository is open source and can be installed from PyPI with pip install agenda-intelligence-md. It also has a pinned wheel available on GitHub for version 0.7.3.

The package ships several components. The Agenda-Intelligence.md protocol gives agents a structured reasoning workflow. JSON schemas validate brief structure, evidence packs, memory cards, lens manifests, and signal trackers. The CLI includes commands such as validate-brief, validate-evidence, score, doctor, bench, source-plan, and mcp-config.

The project also includes a real stdio MCP server, agenda-intelligence-mcp, so MCP-compatible hosts such as Claude Desktop, Cursor, Codex, or custom agents can call validation and scoring tools during an agent loop. According to the project, this avoids copy-pasting between systems and lets the audit layer run alongside generation rather than as a separate step.

The toolkit’s evidence policy is explicit about provenance. Claims can be tagged with categories such as [primary], [secondary], [user-provided], [inference], and [analyst-judgment], plus a freshness marker like [verify] or [stale-risk: YYYY-MM]. The repository also includes a signal lifecycle tracker for moving items through states like detected, developing, escalated, stable, resolved, and archived.

For source handling, the package includes a source-normalization skill that can turn PDF, DOCX, and URL inputs into structured source records for evidence packs. It also bundles regional and sector lenses for Central Asia and the Caspian, the Middle East, the EU, sanctions, and export controls.

The maintainers are clear about the limits. Agenda Intel MD is not a factuality verifier, not a source retriever, and not a replacement for analyst judgment. According to the project, it checks form rather than truth, and the benchmark seed is only a starting point.

The repository says the bundled baseline benchmark covers five cases: EU AI Act, EU CBAM, Red Sea shipping, sanctions routing, and BIS AI Diffusion. It reports a mean score of 87.0 out of 100, 100% schema validity, 100% evidence pack coverage, 100% claim-level audit coverage, and zero orphan evidence references. The project says those results were reproduced with python3 evals/run_benchmark.py.

The CLI is designed for CI-style use. The project shows examples of validating a brief, validating evidence, running claim audits, scoring output with or without an evidence pack, and running a structural benchmark across bundled examples. It also includes doctor to diagnose the local install and MCP tool surface.

In the package structure, the core Python code lives under src/agenda_intelligence/, with schemas, examples, analysis-bank patterns, evaluation assets, docs, and tests alongside it. The project is released under the MIT license.