Glossary

Plain-English definitions of the technical terms used across the bundle.

• Structured output — Constraining a model's output to a schema; guarantees form, not authority.
• Function calling — A model emitting a tool invocation against a declared function schema.
• API calling — Outbound calls to remote services (REST, GraphQL, gRPC, events).
• Tool use — The widest term: APIs, code execution, search, retrieval, workflow calls.
• Tool discovery — Finding that a capability exists (catalogues, agent cards, registries).
• Tool retrieval — Selecting a small candidate set from a large inventory at inference time.
• Schema grounding — Checking that fields and types match a formal contract.
• Semantic grounding — Checking that intent maps to the right operation and entities.
• Agent skill — A reusable, packaged instruction/workflow bundle for an agent.
• API gateway — Runtime enforcement point for authn/z, rate limiting, routing and logging.
• Workflow engine — Orchestrates multi-step, stateful processes; Arazzo describes them declaratively.
• Workload identity — Cryptographic identity for software workloads; 'what is calling'.
• Policy-as-code — Decoupling policy decisions from enforcement points across services.
• Provenance — The recorded entities, activities and agents that produced an outcome.
• Idempotency — An operation that can be safely repeated with the same effect.
• Compensation — Undoing or offsetting a completed step in a multi-step workflow.
• Sender-constrained token — A token usable only by the client that presents the right proof.
• Bearer token — A credential where mere possession grants access.
• Least privilege — Granting only the minimum authority needed for a task.
• Confused deputy — A privileged component tricked into misusing its authority for another.
• Prompt injection — Malicious instructions hidden in tool descriptions or retrieved content.
• Agent Card — A2A's machine-readable description of an agent's capabilities.
• Consent — Explicit user permission for data exposure or action.
• Audit trail — A trustworthy record of what an agent read, decided and did.
• Attestation — Cryptographic proof of a workload's or node's identity/integrity.