Legion Command Papers

Agents are becoming infrastructure. The question is who governs them. These papers examine what changes structurally when AI moves from tool to operator in the environments national security cannot afford to get wrong, and what it takes to deploy, govern, and prove those systems under human authority.

Categories

9 minute read

The AI Cold Start Problem Is a Platform Problem

Time to a first answer is not time to a usable product. Cold start keeps repeating until shared platform layers carry the setup work that users and builders redo for every use case.
12 minute read

Beyond the Answer: Operationalizing AI on Classified Information

An AI agent can draft a sound assessment in seconds and leave the analyst an hour from being able to send it. Why classification markings have to be a system property, and what it takes to preserve them from ingestion through revi
13 minute read

The Intelligence Cycle Breaks at the Seams

Korea 1950, Kuala Lumpur 2000, Curveball 2003. Each time the information existed and never reached the judgment it should have changed. What ICD 505 requires AI to carry across a handoff.
10 minute read

The Denied Column: Military AI for Degraded and Denied Operations

Acetate, cloud, and edge AI ran side by side at Project Convergence. The denied column, the one the brochures skip, is where the war gets decided. Four actions leaders can take before doctrine hardens.
10 minute read

The Kill Chain Is Not a Theory of Victory

Maven compressed the targeting cycle to machine speed. But a faster kill chain wins fights, not wars. The decisive advantage goes to the force that helps commanders turn data into meaning before the decision is made.
16 minute read

Earning a Seat on the Staff

The staff has absorbed new machines for 160 years, each by meeting its standards for coordinated work. What it takes for AI to earn a seat: the five requirements, the three architectural commitments, and how leaders should evaluat
12 minute read

The Forward Deployed Engineering Model Is Backward

Does a vendor's forward deployed engineering model build customer-owned capability or permanent dependency? In defense software, a growing FDE team often signals a product that never matured, not a strength.
7 minute read

Disconnected Is the Baseline: Designing Agentic AI for the Tactical Edge

Most AI architectures treat connectivity as a background condition. At the tactical edge, the assumption fails first. Why disconnected-first design, narrow planners, and sustainment determine if agentic AI works.
11 minute read

Why Retrieval Agents Fail: It’s Not Just the Model

Retrieval-focused agents fail for reasons beyond model capability. Performance is bounded by three independent constraints: the model, the retrieval stack, and prompt guidance. Evaluation finds the real bottleneck.
10 minute read

AI Agent Evaluation: Building an Evaluation Platform That Scales

Teams deploy agents faster than they can test them, and one prompt change can silently degrade three agents while improving one. Immutable datasets, purpose-driven metrics, and vendor-agnostic design make evaluation reproducible.
12 minute read

Agentic AI Security is an Architectural Decision

Agentic AI security is an architectural property, not a post-deployment feature. Governed workflows, strong typing, least-privilege access, and ephemeral agents address 20 OWASP vulnerabilities by design, validated at IL2 to IL6.