Three AWS pieces published the same day (July 13) showcase real agentic implementations in enterprise context. The first tells how Bluesight evolved a prototype on a single product into Prism, a unified agentic system serving six healthcare compliance products. Prism Assistant for ControlCheck has been in production since May 2026 at 20 U.S. health systems, with a more complex multi-product solution arriving by year-end.

The second piece is a complete implementation guide for the on-behalf-of (OBO) token exchange pattern for multi-tenant systems on Amazon Bedrock AgentCore Gateway: it shows how an agent serving many tenants can call downstream APIs while preserving the original user’s identity through JWT audience-bound tokens, without collapsing the audit trail or becoming a confused deputy.

The third post shares the personal experience of an AWS employee who is neurodivergent and uses Amazon Quick on your desktop as a daily accessibility tool, compensating for executive function gaps with an AI assistant that remembers, organizes, and suggests next steps.

Why this matters to you. These aren’t product announcements but documented implementations with architectural details, design choices, and real numbers. Bluesight brought an agent from prototype to production in nine months on HIPAA-eligible data, with session isolation and audit trail; the OBO pattern solves an identity problem that every multi-tenant system must face when an agent calls services on behalf of different users; and the accessibility case shows AI as compensation for invisible disabilities, not as automation of someone else’s work.

If you’re building or evaluating agents in enterprise context, these three cases show three different angles on the same challenge: how you bring an agent to production on real data, with real users, without losing security and observability.

In detail

Bluesight’s path from prototype to production

Bluesight manages compliance for hospitals and pharmacies across the United States. The starting problem was drug diversion detection: compliance teams spent hours filling out reports and correlating signals across multiple dashboards. A conversational interface that could perform this analysis in seconds would save enormous time, but had to respect the security and accuracy standards that hospital compliance programs require.

The second use case was more ambitious: hospitals classified as DSH, PED, or CAN are prohibited from purchasing outpatient drugs through GPO contracts, unless the drug is genuinely unavailable through non-GPO channels. Proving this exception requires evidence from multiple Bluesight products simultaneously: purchase records from CostCheck, shortage data from ShortageCheck, and 340B eligibility from 340BCheck. No single product had the complete picture.

Both cases shared a common constraint: the architecture had to be production-grade from day one. Patient data is governed by HIPAA, compliance teams require audit trails, and any AI system making claims about drug purchase compliance must be explainable and deterministic where it matters.

Bluesight chose AWS because Amazon Bedrock AgentCore provided production-grade agentic infrastructure without having to build it from scratch. Three capabilities proved decisive: Bedrock is HIPAA-eligible (Bluesight operates under a Business Associate Agreement with AWS, and customer data processed by Bedrock is not used to train foundation models); AgentCore Runtime provides secure serverless hosting with session isolation; and the agent-to-agent communication pattern in AgentCore matches the architecture needed for the GPO prohibition case (a coordinator agent delegates to specialized data workers).

From September 2025, Bluesight moved from a ControlCheck prototype to Prism in production across 20 health systems in nine months. The AWS post documents architectural choices, integration patterns, and how AgentCore Gateway transforms existing product APIs into MCP-compatible tools that agents can discover and invoke.

On-behalf-of token exchange: the identity problem in multi-tenant systems

When you deploy AI agents in production multi-tenant architectures, you face a specific identity problem: when an agent calls a downstream API on behalf of a user, whose identity travels with the call? Executing the call as the agent’s service identity collapses the audit trail, because every downstream system must trust the agent unconditionally. Forwarding the unchanged user token turns every downstream tool into a confused deputy. Neither option scales when an agent faces many tenants and the user isn’t present at the moment of the tool call.

The OAuth 2.0 Token Exchange specification (RFC 8693) solves exactly this problem, and Amazon Bedrock AgentCore Identity supports it natively as a credential-provider grant type. The OBO pattern is essential when an agent faces multiple downstream services or tenants and the audience of the inbound token differs from any single downstream API.

The AWS post provides a complete reference implementation (TravelBot, a multi-tenant booking assistant serving two example tenants, Acme and Globex) and shows JWT claim transformations at each hop, demonstrating how audience binding produces defense in depth that scales across tenants. AgentCore Gateway intercepts the tool call, identifies the target tenant, and instructs Identity to execute the exchange against the tenant’s authorization server before the downstream call is issued.

The result is a cryptographically scoped token for a single downstream call on behalf of a single user, with the sub claim preserving the original user’s identity and the actor claim (or cid in Okta) recording who is performing the action. Authorization decisions belong to sub, audit logs and rate-limiting decisions to actor.

AI as an accessibility tool for neurodivergent individuals

The third post is a first-person account of how Amazon Quick on your desktop serves as a daily accessibility tool for a neurodivergent professional. The system compensates for executive function gaps that make it difficult to track tasks, priorities, and next steps in a complex work environment.

It’s not a story about automating someone else’s work, but about compensating for invisible disabilities: the AI remembers what you were doing, suggests the next step, and organizes information in a format your brain can process. It’s the accessibility angle that rarely appears in enterprise AI announcements, but that matters for a significant fraction of the workforce.

This use case is also an example of how AI enters the enterprise not as a replacement for people but as a tool that enables people to work better. This site’s manifesto says AI serves to empower people, not replace them: this is a concrete case where that distinction is clear and measurable.