From the AWS edition of October 29, 2026
AWS outlines a human-in-the-loop autonomy model for agentic SNAP and Medicaid eligibility processing on Bedrock AgentCore
Amazon Web Services (AWS Public Sector Blog) · Health and human services (state eligibility programs) · United States
- Publisher
- AWS Public Sector Blog
- Original publication
- September 1, 2026
- Source retrieved
- 2026-10-30
What happened
AWS argues most production eligibility agents should not start fully autonomous, noting states face sanctions when SNAP determination error rates exceed 6%. It proposes three modes: Assist (agent drafts, caseworker approves each decision), Supervised autonomy (agent handles routine cases; edge cases, denials and low-confidence cases go to humans) and selective Full autonomy for case types with sustained accuracy. Cedar policies declaratively define what the agent may do, per program, such as authorizing SNAP determinations while keeping Medicaid recommend-only. Knowledge Bases, AgentCore Memory, Runtime, Gateway and Guardrails provide policy retrieval, case context, isolation, tool access and PII protection.
Why it matters
Directly applicable to state and county HHS agencies weighing agentic AI in eligibility, where error rates carry federal financial consequences.
Evidence and measured results
Vendor architecture guidance; no named state deployment or measured accuracy results. The SNAP 6% threshold is cited by AWS.
Limitations and uncertainty
Vendor guidance, published about two months before this edition; no field results; federal program rules and state policy determine what any autonomy level can legally do.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-29; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Eligibility caseworkers face backlogs while error rates carry federal financial consequences.
- Stakeholders
- State HHS secretary or deputy
- SNAP and Medicaid program directors
- Eligibility operations managers
- Quality control staff
- Agency counsel and CISO
- Discovery questions
- What is your current payment and determination error rate by program?
- Which case types are most routine?
- Who must approve changes to agent authority?
- Which systems would the agent need to call?
- Value hypothesis
- An Assist-mode agent may shorten caseworker drafting time while preserving accuracy, to be validated against quality-control review.
- Potential engagement
- An Assist-mode pilot on a narrow case type with accuracy baselining and policy-as-code design.
- Unsupported claims
- Accuracy or error-rate improvements
- Any state operating autonomous determinations
- Federal approval of autonomy levels
Pre-sales engineering
Role takeaway
- Fit
- Fits agentic drafting with human approval on Bedrock AgentCore.
- Architecture and integration
- AgentCore Runtime, Memory and Gateway; Knowledge Bases for policy; Cedar for authorization; integration with case management and verification hubs.
- Prerequisites
- Policy manuals in retrievable form
- Case management API access
- Quality-control baseline data
- Constraints
- Federal program rules
- Data-sharing agreements for verification hubs
- Denials must remain human
- Security
- Cedar caseload scoping, Guardrails for PII and prompt injection, full decision audit logs.
- Proof of value
- Draft accuracy versus caseworker decisions
- Zero policy-violating agent actions in testing
- Time saved per case in Assist mode
Delivery
Role takeaway
- Work
- Deploy an Assist-mode agent on one routine case type and measure accuracy.
- Dependencies
- Program policy staff
- Eligibility system vendor
- Security and privacy office
- Ownership
- State HHS eligibility operations with IT.
- Skills and adoption
- Train caseworkers to review agent drafts and log disagreements.
- Governance checkpoints
- Policy-as-code review by program and legal staff
- Accuracy gate before any autonomy expansion
- Periodic error-rate reporting
- Acceptance criteria
- Proposed: agent drafts match final determinations at an agreed rate
- Proposed: all denials human-decided
- Proposed: auditable decision trail for every case
- Risks
- Erroneous determinations harming families
- Automation bias among reviewers
- Federal compliance exposure
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Separates authorization (Cedar policies) from agent logic; routes tools through AgentCore Gateway to case management, wage databases and federal verification hubs; uses isolated runtime and session memory.
Governance
Who approves, reviews and stays accountable for outcomes?
Autonomy expands by evidence of measured accuracy, with versioned, auditable policies; denials and edge cases stay with humans.
Security and privacy
What data, permissions and controls need testing?
Cedar limits case access to caseloads; Guardrails aim to prevent PII leakage and prompt injection; access to federal data hubs carries its own compliance obligations.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Caseworkers shift from drafting to reviewing and handling exceptions; training on reviewing agent drafts is needed.
Procurement
What should contracts, pricing and exit terms secure?
Statements of work should tie autonomy levels to measured accuracy gates and require policy-as-code artifacts the state owns.
Operating model
Which teams own the service once it runs?
Requires ongoing measurement of error rates by case type and a formal process for expanding or retracting agent authority.
Publication history
- 2026-10-29AWS · Issue 019 resources
Stable resource ID: aws-agentcore-human-in-the-loop-snap-medicaid-eligibility-20260901