Agentic AI on AWS · built for regulated workloads

AI agents your compliance team can actually sign off.

Most agent demos put a model in front of your data and hope. This is the other thing: a deployed agent platform where every tool call is authorised before it runs, memory is encrypted per-tenant, and the whole stack is Terraform you can read in a pull request.

Amazon Bedrock AgentCore Cedar policy MCP Terraform HIPAA-eligible services Azure — in progress

The mechanism

One request, four chokepoints it cannot route around

The interesting question about an agent is not what it can do. It is what it cannot do, and who decided. Follow a single tool call:

denied — tool never runs AI client MCP Identity SigV4 / OIDC Cedar deny by default Gateway tool catalogue Tool λ CloudTrail · every hop recorded, including the denials
The policy decision happens before the tool executes, in the request path rather than in the prompt — so a jailbroken or simply confused model cannot argue its way past it. A denied call never reaches your data.

Governance

The only question worth asking about a control

Not “do we have it?” but can the agent bypass it? Three honest answers, and knowing which is which is most of the job:

Hard — no way around

Infrastructure

VPC-only networking with an egress allowlist, the managed runtime itself, Cedar at the gateway, organisation CloudTrail. The agent has no path around these because they are not in its control.

Conditional

Only on the happy path

Model guardrails are real if every call goes through the gateway — and silently bypassed by a direct SDK call. Most teams believe these are hard controls. They are not.

Convention

People, not infrastructure

Prompt management, orchestration style, evaluation discipline. A template can seed them; nothing enforces them. Pretending otherwise is how audits go badly.

PHI workflow

Dev to prod, one click, nothing manual in between

Regulated work fails on process far more often than on architecture. Three environments, identical infrastructure code, promotion by merge — and a human approval gate that exists in infrastructure rather than in a runbook.

Dev synthetic data only — no PHI Stage de-identified Production PHI · BAA in place · CMK scan · plan · policy runs on every pull request smoke test · drift check fails the merge, not the customer named human approval enforced by infrastructure
Promotion is a merge, not a person running commands. The approval gate is a protected environment rather than a policy document, so it cannot be skipped when someone is in a hurry — which is precisely when it matters.
Said plainly

AWS services can be HIPAA-eligible; architectures are not “HIPAA certified”. What this buys you is a Business Associate Agreement in force, PHI confined to services covered by it, encryption under keys you control and can revoke, US-region inference binding, and an audit trail that survives contact with an assessor. The controls are real. The certificate does not exist — and anyone selling you one is worth walking away from.

The decision space

Every layer is a choice, and the wrong default is expensive

None of this is one product with one shape. Each layer has real alternatives, and picking badly early is what makes agent projects stall in review. Highlighted is what a regulated healthcare deployment usually wants — not because it is the only answer, but because it is the one that survives an assessor.

Inbound auth

who may call at all
IAMCognito JWTEntraOktaAuth0none

Outbound auth

how the gateway reaches the tool
gateway IAM roleAPI key vaultOAuthJWT passthroughcaller identity

Tool source

what is actually behind it
LambdaOpenAPI specexisting MCP serverAPI Gatewayknowledge baseweb search

Policy

what may run, and when
Cedar, deny by defaultcontent filtersprompt-attackPII suppressionlog-only first

Agent host

where the loop runs
config-only harnessruntime in your VPClocal, for proving it offline

Model access

cost attribution and residency
tagged inference profilebare model idOpenAI / Gemini via vaultself-hosted

Memory

what the agent remembers, and whose
none, for public surfacesper-tenant with your own keymanagedsemanticpreferencesepisodic

Network

the difference between hard and hopeful
VPC-only + egress allowlistpublic

The two that catch people out: memory defaults to a shared pool unless every caller is scoped, so one visitor’s context can surface in another’s answer — and encryption keys cannot be changed after the memory exists, so “we will add the customer key later” means rebuilding it.

Getting started

How this usually goes

Nobody should begin with a platform. Begin with one workflow that is expensive, repetitive, and currently done by a person reading three systems at once.

  1. A two-week readiness assessment

    Current-state review of your AWS or Azure footprint against the controls that agent workloads actually need: network egress, model access scoping, key management, logging retention, and where PHI would flow. You get a gap register and a prioritised roadmap. Fixed fee, no platform commitment.

  2. One agent, one workflow, non-production

    Built on synthetic data, deployed through the same pipeline production will use. The point is not the demo — it is proving the promotion path, the approval gate, and the audit trail before any real data exists.

  3. Promote, with the guardrails already in place

    Cost caps, policy in enforce mode, per-tenant key isolation and alerting are provisioned before the first real request, not after the first surprise invoice. Then the second workflow costs a fraction of the first.

Evidence

Numbers from the running system, not a proposal

Everything on this page describes a platform that is deployed and serving requests today. These are measured figures, not estimates.

$0.0016per question End to end, including tools and memory. 700 ms median.
100%Terraform 21 AgentCore resource types, one state file, one destroy.
4clouds, live AWS, Azure, GCP and OCI serving the same site with DNS failover.
25governed tools Published once, authorised individually, versioned in git.

The part worth trusting is not that it works. It is knowing the ways it can appear to work while being wrong — a tool schema that silently drops every argument so the model invents them; a smoke test using curl without -f, which exits zero on an HTTP 500 and reports a broken deploy as green; a certificate renewed correctly while the gateway keeps serving the expired version it pinned months ago. Each of those was found and fixed here, and each is now a guard that fails loudly.

Next step

Worth a conversation if any of this sounds familiar

You have an agent proof-of-concept that security will not let near production. Or a compliance deadline and a platform nobody can explain in a pull request. Or you simply want to know what this would cost before committing to anything.

Fifteen years in healthcare IT, HIPAA and SOC 2 environments, AWS landing zones and Terraform — and the agent platform described on this page is one I built and operate, not a slide.

Start a conversation See a live agent