Glossary

Canonical terms used throughout the Agent Stream documentation.

Agent Stream is the model-governance layer for AI agents: it sits between callers and the LLM providers, APIs, and tools they call, and applies routing, guardrails, budgets, and observability to every execution from one configuration layer. It is part of the Affinidi Trust Fabric.

This glossary provides clear, consistent terminology to help you get started quickly.

Routing model

Term	Meaning
Surface	The single configuration and runtime unit Agent Stream uses to represent one governed connection. Surfaces →
LLM Surface	A surface configured against one LLM provider. Carries the full guardrail, routing, resilience, and cost pipeline.
IDE Surface	A governed, per-caller-filtered model menu that aggregates one or more LLM Surfaces behind a single OpenAI-compatible endpoint, for editors and OpenAI-compatible agents. IDE Surface →
Variant	A named overlay on a surface, addressable at `{route}$alias/...`, that overrides part or all of the surface’s behaviour (provider, model, guardrails, limits) without a separate surface record. Variants and progressive rollout →
Access Point node	The canvas node (labeled Access Point) on any surface, LLM or IDE, that defines how incoming requests reach it: listen address, path prefix, and custom path, composed into the surface’s route. Distinct from a connection point, which is the inbound DIDComm endpoint used for gateway-to-gateway traffic.
LLM node	The canvas node (labeled LLM) on an LLM Surface that defines the provider, model, and API key secret the surface calls out to.
Route	The composed full URL shown on a surface’s Access Point node: listen address, path prefix, and custom path combined. Callers send requests to this address.

Providers and wire formats

Term	Meaning
Provider	The LLM backend an LLM Surface reaches: OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure OpenAI, Microsoft 365 Copilot, Ollama, Groq, xAI (Grok), OpenRouter, Fireworks AI, Cerebras, DeepSeek, Moonshot AI (Kimi), SambaNova Cloud, Hyperbolic, NVIDIA NIM, Nebius AI Studio, or the built-in Marketplace.
Model catalogue	The hot-reloadable source of provider pricing, capabilities, and request profiles. A new model becomes available through the dashboard, not a code change.
Wire format	The request/response dialect a client uses to call a surface: OpenAI Chat Completions, Anthropic Messages, or OpenAI Responses. Independent of which provider the surface itself is configured to reach.
Starter template	A pre-built canvas layout offered when creating a surface, such as OpenAI Chat Completions or Blank LLM Surface, that pre-wires a provider, model, and endpoint.
Partial template	A template that merges a bundle of elements onto an existing surface instead of replacing it, such as the Guardrails Pack.

Guardrails

Term	Meaning
Guardrails	The layered content-safety pipeline applied to a surface’s traffic. Guardrails →
Prompt Guard	Deterministic pattern matching plus model-backed AI detection for PII, run on the request (both) and response (pattern only) seams. Each match can be rejected, redacted, or replaced with a NER ID.
NER ID	A stable pseudonym assigned to a detected PII value, preserving relationships between values without exposing the raw data. PII protection and NER IDs →
Expert Witness	A purpose-trained external safety classifier (for example EnkryptAI, Lakera Guard, Azure Prompt Shields, Azure Content Safety) consulted on the request seam, the response seam, or both.
Judge	An LLM-based or moderation-based pre-check that runs on the request in parallel with routing, and can block a request before the main model is called.
Jury	A multi-juror LLM post-review of the model’s response that can approve, block, or regenerate an answer, with optional per-juror retry.
Fail closed	The default guardrail behaviour: if a guardrail provider errors or times out, the request is blocked rather than allowed through unchecked.
Rejection taxonomy	The stable set of categories (hate, toxicity, prompt injection, jailbreak, PII, secrets, policy violation) every guardrail block is classified into, charted per team, member, surface, and stage.

Access control

Term	Meaning
OPA (Open Policy Agent)	The embedded Rego policy engine, evaluated in-process via regorus with no external OPA server required. OPA policies →
Gateway-level policy	Policy evaluated first, before any surface-level policy; a gateway-level deny is final.
Surface-level policy	Policy attached to one surface, evaluated after gateway-level policy.
RBAC role ladder	The three roles governing dashboard and API access: administrator (every permission), poweruser (adds mediators, connection points, and surface capture on top of the user role, still with no secrets or user/RBAC management), and user (the default role, with broad view access plus edit and delete on LLM and IDE Surfaces, payments, notifications, and departments; no secrets, user/RBAC, or settings access). Security and access control →
Secrets store	The encrypted store holding provider API keys and other credentials, referenced by ID from a surface rather than embedded in configuration.
Source authentication	Verification of the inbound caller’s identity: JWT bearer token, API key, or mTLS.

Cost, attribution, and monitoring

Term	Meaning
Usage limit	A hard ceiling, per surface, per stage, or per juror, on monthly USD spend or monthly token usage. A request that would exceed it is blocked with `429` before it reaches the provider.
Breach alert	A notification, not a block: fires after a cost or token threshold is crossed, or an anomaly is detected, so the request still completes.
Member	A registered caller, a person or a client, that reaches models through the appliance.
Team	A group of members. Usage is attributed to both the member and its team on any surface where attribution is enabled.
Virtual key	The effective behaviour of a member: a model allow-list, per-minute rate limits, and monthly ceilings, on top of a surface’s own limits. Teams and attribution →
x-agent-stream- headers*	Per-request response headers carrying trace ID, provider, model, token counts, cached-token count, cost, cache hit or miss, and ingress wire format.
Governance record	A signed W3C Verifiable Credential produced per interaction, when recording is enabled, capturing the prompt, every stage’s verdict, the completion, and metrics, chained to the previous record on the same surface. Governance records →

Resilience

Term	Meaning
Failover	Routing to ordered backup providers when the primary is unreachable or erroring.
Load balancing	Distributing traffic across a pool of providers by weighted round-robin, lowest observed latency, or lowest catalogue cost.
Response cache	An in-memory cache serving a repeat or near-duplicate prompt without calling the provider, in exact (hash) or semantic (embedding-similarity) mode. Not available for streamed responses.
Circuit breaker	A control that stops forwarding requests to a failing provider after a threshold of failures, resetting automatically after a cooldown.

Trust fabric

Term	Meaning
G2G (gateway-to-gateway)	DIDComm v2.1 traffic between Agent Stream instances, addressed as `fabric://{gateway_id}/{channel_id}`.
Connection point	The inbound DIDComm endpoint one appliance exposes for another to reach it, over G2G.
Mediator	A component that routes DIDComm messages between instances that cannot connect directly.
x402	An HTTP 402 payments protocol for blockchain-settled, metered access to a model or service.
Model marketplace	A built-in catalogue and purchase flow letting a caller provision an API key for a purchased model directly from the appliance.

Was this page helpful?

Glad to hear it! Please tell us how we can improve more.

Sorry to hear that. Please tell us how we can improve.

Thank you for sharing your feedback so we can improve your experience.