
Governing agent access with AgentCore Gateway and Okta grants
The post describes an architecture in which Amazon Bedrock AgentCore Gateway is the single front door for agent tools. For permitted requests, the Gateway uses AgentCore Identity to redeem an Okta grant for the relevant target. The front door therefore represents more than a shared endpoint: policy evaluation, credential acquisition, and target invocation form a defined sequence.
A decision before target invocation
AgentCore Policy evaluates Cedar policies for every tool listing and every tool call before the Gateway invokes a target. When Cedar permits a call, the Gateway uses AgentCore Identity to redeem the ID-JAG for that target. When Cedar denies the call, it does not redeem the ID-JAG. This ordering separates the question of whether an agent may use a tool from the acquisition of the credential needed to make an authorized request.
The post assigns two related but distinct authorization responsibilities. Cedar binds tools to an agent and to the agent's delegated scope. The MCP server determines the authority of the individual user from the groups claim. In this design, an agent's access to a tool and a user's authority at the target are not treated as the same decision.
One identity boundary per agent
Each agent in the described configuration receives its own Gateway and credential provider. The stated reason is that a credential provider authenticates as one Okta agent. Keeping these resources separate preserves the distinction between agent identities while allowing them to follow the same overall governance pattern.
The second, reporting agent runs on Amazon EKS. It has its own Okta registration, Gateway, and credential provider, and its resource connection permits only xaa:read. The example applies the same architectural pattern to an agent with a narrower role without merging its identity or delegated scope with those of another agent.
Validation outside AgentCore Runtime
Because the EKS-hosted agent operates outside AgentCore Runtime, it validates the ID-JAG before making a model call. The checks described in the post cover the signature, issuer, audience, token type, expiry, and client_id. These are explicit duties in the presented configuration; they should not be read as evidence that every externally hosted agent performs them automatically.
Tracing the path and recording grants
Runtime, Gateway, Identity, and Policy emit OpenTelemetry spans for each hop after Transaction Search and trace delivery are enabled. Custom services add their own spans with the AWS Distro for OpenTelemetry. The resulting account of the request path can cover policy evaluation, identity handling, and progression toward the target.
A separate audit path begins with Okta System Log. Okta can deliver its events to an Amazon EventBridge partner event bus, where a rule writes them to a CloudWatch Logs log group. The post therefore places operational traces beside a stream of identity-related events rather than presenting either record as the sole view.
In the reported tests, Okta System Log events arrived about one second after publication. Each ID-JAG grant event identified the agent, the user, the granted scopes, and the resource connection. This is a test-specific observation. It does not establish a guaranteed delivery time for other configurations or deployments.
What the deployed-agent checks showed
The reported checks showed agent A denying Alice and Carol while creating Bob's task. Agent B told Bob that it had no tool for creating tasks. The contrast demonstrates the described separation between user authority and the tools bound to a particular agent: a user's position does not by itself make the same tool available through every agent.
The result also reflects two different denial points. A request may be excluded by the agent's available tools and delegated scope, while the MCP server separately evaluates the user's authority using the groups claim. The evidence reports the observed outcomes but does not turn those outcomes into a universal statement about all agents or all policy sets.
Where XAA fits
According to the post, XAA is used when a target authorization server accepts an ID-JAG. Applications that have their own authorization servers, with GitHub given as an example, instead use each user's OAuth consent. The architecture is therefore not presented as a claim that every possible target can or should use the same grant mechanism.
Taken together, the design combines a single Gateway front door, pre-invocation Cedar decisions, per-agent identity resources, target-specific grant redemption, user authorization at the MCP server, and visibility across traces and identity events. The supplied evidence describes conceptual sample code and configurations rather than a complete production-ready deployment. Its timing, behavior, and interoperability observations remain results of the reported tests, not guarantees of availability, performance, security, compliance, or scalability.