6.1. Terms A–Z

A2A (Agent-to-Agent) protocol — Open, cross-vendor protocol (Linux Foundation) for peer-agent collaboration via agent cards and managed tasks. (1.3.2, 2.3.4)

Adaptive card — JSON-defined interactive UI in a message: inputs and actions whose submitted values return to topic variables. (3.3.2)

Agent card — Machine-readable capability description an A2A agent publishes for discovery. (1.3.2)

Agent flow — Deterministic automation authored, run, and billed inside Copilot Studio; invocable by agents via the "When an agent calls the flow" trigger. (3.2)

Application Insights — Azure telemetry service; receives agent telemetry after its connection string is configured; queried with KQL. (2.4.3)

Approval (human-in-the-loop) — Flow action that routes a decision to named approvers, waits statefully, and branches on the verdict. (3.2.2)

Azure AI Search (knowledge source) — Bring-your-own vector index grounding; requires integrated vectorization; one index per connection; URL field for citations. (2.1.3)

Child agent — Lightweight subagent inside a parent agent, sharing its configuration and auth; groups tools/intents for routing accuracy. (2.3.1)

Compare Meaning — Evaluation method scoring semantic similarity (0–100%) against an expected answer. (4.1.2)

Computer use — Tool that operates website/desktop UIs on a dedicated Windows machine via a vision-reasoning model; requires generative orchestration; billed per step. (2.2.1)

Connected agent — Independent agent (Copilot Studio, Foundry, Fabric, or A2A) that a parent delegates to; own lifecycle and settings. (2.3)

Connection reference — Solution component letting each environment rebind connector credentials without editing components. (4.2.1)

Copilot connectors — Formerly Graph connectors; ingest external content into the Microsoft 365 semantic index as external items with ACLs. (2.1.1)

Copilot Credits / capacity — Consumption billing for agent and flow usage; exhausted capacity blocks new flow runs. (3.2.1)

Custom connector — OpenAPI-defined wrapper making your API a governed, reusable Power Platform connector; republish to change. (2.2.3)

Custom prompt — AI Builder prompt (instructions + inputs + model) used as a tool or topic node; can select models from the Foundry catalog. (2.4.2, 3.3.3)

Dataverse — Power Platform's data backbone; common structured knowledge source and solution store. (1.4.1)

DLP policy — Data loss prevention rules classifying connectors and blocking disallowed combinations across the environment. (3.1.3)

Environment variable — Per-environment configuration slot (text, data source, or secret) supplied at deployment. (4.2.1)

Exact Match — Pass/fail evaluation requiring character-for-character equality with the expected answer. (4.1.2)

Fabric data agent — Microsoft Fabric agent answering natural-language questions over lakehouses/warehouses/KQL; orchestrator-routed only; not compatible with M365 Copilot deployment. (2.3.3)

Foundry (Microsoft Foundry) — Azure's AI platform (rebranded from Azure AI Foundry): model catalog, fine-tuning, pro-code Foundry agents. (1.4.2)

General Quality — Evaluation method judging relevance, groundedness, completeness, and abstention without an expected answer. (4.1.2)

Generative answers node — Topic node performing scoped RAG: per-node data sources, custom instructions, moderation level, citations; node sources override agent knowledge. (3.3.4)

Generative orchestration — Mode where the model routes among topics, tools, knowledge, and agents based on their descriptions. (1.1.2)

Keyword Match — Pass/fail evaluation checking for required words/phrases (any or all). (4.1.2)

KQL — Kusto Query Language; queries agent telemetry in Application Insights. (2.4.3)

Managed / unmanaged solution — Sealed deployment artifact vs. editable workbench; unmanaged in dev, managed downstream. (4.2.1)

MCP (Model Context Protocol) — Open protocol connecting an agent to a server's tools/resources with run-time discovery; server-side additions propagate without republishing. (1.3.1, 2.2.2)

Message variations — Alternate phrasings on a message node to avoid robotic repetition. (3.3.2)

Pipeline (Power Platform Pipelines) — Staged, auditable solution promotion (dev → test → prod); extensible with approvals/validations; supports service principals. (4.2.2)

Power Fx — Power Platform formula language; reads/transforms variables in topics. (3.3.5)

Run-after condition — Per-action setting routing on succeeded/failed/skipped/timed-out; the basis of flow error handling. (3.2.4)

Send HTTP request node — Topic node making raw API calls and parsing responses into variables. (3.3.3)

Service principal — Non-human identity for unattended operations such as automated pipeline deployments. (4.2.2)

Solution — Versioned package carrying an agent and all dependencies between environments. (4.2.1)

System topics — Platform-provided topics (Conversation Start, Fallback, Escalate) handling conversational events. (3.3.1)

Test set — Batch of test cases (input + success criteria) run repeatably against the agent with chosen evaluation methods. (4.1.1)

Text Similarity — Evaluation method scoring wording/structure closeness (0–100%) to an expected answer. (4.1.2)

Tool Use — Pass/fail evaluation verifying specific tools/topics were used to produce the response. (4.1.2)

Topic — Authored conversation unit: selection face (description/trigger phrases) plus deterministic node canvas. (3.3.1)

Variable scope — Topic (default), global (promoted, conversation-wide), or system (platform-set). (3.3.5)