Public-sector agentic AI deployment in 2026 operates under five constraints that materially narrow the vendor and architectural options compared to private-sector deployment: (1) FedRAMP authorisation (Moderate or High depending on data sensitivity) is required for federal deployments and increasingly for state, (2) sovereign data residency requirements (data and model inference must remain within national or sub-national boundaries), (3) procurement transparency obligations (the deployment, the vendor, and the decision logic typically must be publicly disclosed), (4) explicit accountability under administrative law (decisions affecting individuals are subject to due-process and appeal frameworks that the agent must support), (5) FOIA-equivalent disclosure of audit logs to the public on request. Public-sector deployments cannot reasonably use peer-to-peer multi-agent patterns and cannot accept vendors without published government cloud SKUs; the realistic 2026 options are Microsoft Azure Government, AWS GovCloud-deployed Anthropic, Google Cloud Public Sector, and a small number of specialist government-AI vendors. The NYC MyCity case (claim AM-044) is the canonical 2026 public-sector failure illustrating what happens when the constraints are inadequately addressed.

Public-sector agentic AI procurement constraints. 90-day review cadence. Watches: (1) the OMB M-24-10 successor framework (post-Executive-Order-14110 federal AI guidance is actively evolving), (2) FedRAMP framework updates including the AI-specific authorisation provisions in development, (3) state-level AI procurement laws (Colorado, Utah, Texas, California, Washington) that establish state-specific procurement bars, (4) the NIST AI Safety Institute's outputs that increasingly serve as de facto federal procurement criteria, (5) emerging case-law on public-sector AI deployment liability.