A2A and API Design: 5-Question Check

Explanation: An Agent Card tells a client what a server advertises and how it expects to be contacted. It is useful discovery input, but a sensitive task still needs the endpoint, authentication scheme, authorization boundary, and deployment trust policy to be checked. A: Reachability does not independently prove every identity claim, skill, or permission in the document. B: This keeps discovery separate from the security checks required before disclosing data or invoking a capability. C: HTTPS protects a connection to the authenticated server name; it does not authorize every advertised operation. D: The A2A 1.0 specification makes Agent Card signatures optional and recommends verifying them when present. An unsigned Card still needs a trusted delivery path and a local policy decision; absence of a signature alone is not a universal rejection rule.

Explanation: Authentication and token validity are inputs to an authorization decision, not a substitute for it. A2A requires authorization checks on operations, while OAuth scopes and resource boundaries limit what a token can be used to access. A: Advertising a skill describes capability; it does not grant every caller permission to invoke it. B: Expiration is only one token check and does not establish access to the requested resource or action. C: This evaluates the credential in the context of the exact caller, operation, resource, and server policy. D: An unverified client assertion does not replace a server-side authorization check.

Explanation: A lost response creates an unknown outcome: the server might already have accepted the task. RFC 9110 warns against automatically retrying non-idempotent requests unless the client knows the operation is idempotent or can determine that the original request was not applied. A: Blind repetition can duplicate purchases, messages, or other side effects. B: Creating a new task can still duplicate the accepted work and loses correlation with the first attempt. C: Retries are possible when the method or application contract supplies a safe recovery rule. D: Status lookup and defined deduplication make the uncertain result manageable. A2A 1.0 specifically allows an agent to use messageId to detect duplicate Send Message operations.

Explanation: Useful diagnostics can usually correlate events with task, request, error, and timing identifiers. OAuth credentials must remain confidential, and privacy guidance favors data minimization, limited retention, and access appropriate to the operational purpose. A: Tokens and full prompts create credential and personal-data exposure that ordinary correlation does not require. B: This preserves operational traceability while reducing the amount of sensitive material available through logs. C: Reversible content remains sensitive and can be exposed when the application or key is compromised. D: Short retention does not remove the need for access controls or data minimization during that period.

Explanation: TLS 1.3 is designed to provide an authenticated, confidential, integrity-protected channel. The application still has to decide whether the authenticated endpoint, caller, requested action, data, and current policy fit together. A: Channel authentication does not grant business-level authority for arbitrary actions. B: TLS protects transport to the named server but does not validate every application claim or its freshness. C: This states the channel guarantee without extending it into identity, capability, or authorization claims above TLS. D: Application-level duplicate handling and replay consequences still need their own protocol rules.