AWS Added OAuth Discovery. Don’t Fix a Metadata Error by Rotating Secrets. AWS added OAuth 2.0 discovery URL support to Amazon Bedrock AgentCore Identity on November 13, covering both inbound JWT authorization and outbound OAuth credential providers. The change lets AgentCore work with authorization servers that publish OAuth metadata under RFC 8414's well-known path without OpenID Connect, where the well-known segment sits between host and issuer path rather than being appended as in OIDC discovery. The source notes that as of November 15 the inbound authorizer guide still described an OIDC-only discovery URL pattern, and advises verifying metadata retrieval before rotating client secrets. A provider setup fails, and the first instinct is to replace the client secret. That can create a second problem while leaving the first untouched. On November 13, AWS added OAuth 2.0 discovery URL support to Amazon Bedrock AgentCore Identity https://aws.amazon.com/about-aws/whats-new/2026/11/bedrock-agentcore-oauth2-discovery/ . The practical payoff is compatibility with authorization servers that publish OAuth metadata without OpenID Connect. Before changing credentials, establish whether the integration ever retrieved the right metadata. AI disclosure: This article was generated with AI assistance from the public sources linked below, with automated fact-checking and editorial inspection. It has not been reviewed by Keith. AWS says the update covers inbound JWT authorization and outbound OAuth credential providers. Previously, AgentCore accepted only OIDC discovery URLs, excluding OAuth-only servers that publish metadata at a different well-known path. That is a specific compatibility change, not a promise that every OAuth integration will work without further configuration. The distinction becomes easy to miss when an issuer includes a path. Consider this fictional issuer: https://identity.example/teams Under OIDC Discovery, its metadata location follows this shape: https://identity.example/teams/.well-known/openid-configuration For OAuth Authorization Server Metadata using the default RFC 8414 suffix, the shape is: https://identity.example/.well-known/oauth-authorization-server/teams Notice where teams moved. RFC 8414, published in June 2018 https://www.rfc-editor.org/rfc/rfc8414.html section-3.1 , inserts the well-known segment between the host and issuer path; its compatibility notes explain the difference from OIDC’s appended form. Section 3.3 also requires the returned issuer to match the issuer used to construct the metadata request. Those examples illustrate standards behavior, not a report of a failed AgentCore deployment. Use the provider’s documented discovery URL rather than building one by replacing a suffix in an old configuration. A root-level issuer can conceal the path-order difference during a quick test. For an outbound integration, I would make the setup record show the issuer, discovery URL, selected client authentication method, and last successful configuration check. These are proposed operational fields, not claims about AgentCore’s console interface. Keep secrets and tokens out of that record. Then investigate the observed failure in order. Could the service reach the metadata document? Did it receive usable metadata for the intended issuer? Did the token exchange fail after discovery succeeded? Did the downstream resource reject the resulting request? Each answer points to a different owner and a different corrective action. Rotating a secret belongs to credential troubleshooting when the evidence supports it. It cannot make an incorrectly constructed metadata URL resolve. Likewise, a successful metadata fetch does not demonstrate that a particular member can access a particular application record. The current custom-provider guide https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/identity-add-oauth-client-custom.html describes discovery as automatic endpoint retrieval and also offers manual configuration. It still requires a client ID and client authentication configuration. Discovery reduces configuration work; it does not supply an application’s credentials. There is also a documentation wrinkle worth recording in a rollout ticket. As checked on November 15, the inbound authorizer guide https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/inbound-jwt-authorizer.html still describes an OIDC-only discovery URL pattern, while the announcement explicitly includes inbound OAuth discovery support. Verify the configuration path you plan to deploy. Do not quietly turn that inconsistency into a copied implementation assumption. The relevant connection to my work is the separation of information flows. The AfroTech Guide case study https://asaday.co/projects/afrotech-directory documents an independent community-built public guide and connected AI experience. Approved public records stay separate from private account state and contact fields. Its public website and connected features have different release scopes; real-device and AI-host acceptance remain separate checks. That architecture makes a useful troubleshooting example. Suppose a conversational planning service can search approved public events, while saving a member’s plan requires account access. A discovery configuration failure should be distinguishable from an application refusing to save someone else’s plan. They may surface in the same chat window, but combining them into “authentication failed” makes support harder. This is an illustrative extension of the documented design, not a claim that the guide uses AgentCore. The lesson is to carry the failed operation through to the support view. “Could not load provider metadata” and “save denied for this account” should lead to different investigations. For a team adopting the new support, the useful deliverable is a small, repeatable provider onboarding check. Record the provider-documented URL and expected issuer, verify discovery through the actual deployment’s network path, then exercise the intended account operation with synthetic records. Include a rejection case: a valid signed-in account should still fail when it attempts an operation outside its application access. That keeps a compatibility fix from being mistaken for a completed authorization review. Manual endpoint configuration can be a practical alternative when discovery is unavailable, but it creates configuration that somebody must maintain as the provider changes. Choose it deliberately and name that owner. AWS removed a real onboarding blocker. The next improvement is operational: give the person handling a failed connection enough evidence to fix the right layer before changing a working credential.