| A2A | Agent-to-Agent protocol, donated by Google to the Linux Foundation, for cross-vendor agent coordination and trust. | Stop deploying agents, start onboarding them |
| ACL | Access Control List, the per-item permission record that retrieval layers honor when surfacing content to a user. | Dataverse is the agent backend most M365 shops already own, SharePoint is the grounded RAG layer you already own, Microsoft 365 ships agent inventory not observability |
| ACS | Agent Control Specification, an open middleware layer between agent runtime and a policy engine using Rego or YAML. | Observability outlives your agent framework |
| Agent 365 | Microsoft control plane for observing, governing, and securing AI agents across the tenant, priced per user. | Stop buying Copilot seats, rebuild one process instead, Governance is the new agent bottleneck, Stop deploying agents, start onboarding them, Microsoft 365 ships agent inventory not observability |
| Agent Flows | Copilot Studio workflow surface with LLM-native nodes (Prompt, Classify, Agent) sitting next to Power Automate connectors. | Boring ML on Fabric still beats LLM-everything for the Mittelstand, Stop buying Copilot seats, rebuild one process instead, Agent Flows are the SMB onramp to agentic automation, Copilot Studio Workflows is the spine LLM agents needed, Models depreciate, eval suites compound |
| agent loop | The plan, call-tool, observe, decide cycle a model runs inside an orchestrator, typically with checkpointing. | Mittelstand AI runs on a Mac Studio, not someone else's cloud, Boring ML on Fabric still beats LLM-everything for the Mittelstand, Stop routing every agent step through a frontier LLM, Agent Flows are the SMB onramp to agentic automation, Structured retrieval beats vector RAG for enterprise agents |
| agent skill | A markdown SKILL.md file with frontmatter that tells a coding assistant when and how to follow a team workflow. | Agent Skills are the cheapest ALM your team will ever buy |
| agentic | Software that reasons over tools and state at runtime rather than executing a fixed program. | Pick the durable runtime before your second agent ships, Stop routing every agent step through a frontier LLM, Citizen-dev Copilot agents still need pro-dev ALM, Memory tiers, not bigger models, cut your agent token bill, Agent Flows are the SMB onramp to agentic automation, Four agent patterns recur regardless of vendor stack, Copilot Studio Workflows is the spine LLM agents needed, Structured retrieval beats vector RAG for enterprise agents, Models depreciate, eval suites compound |
| Agentic CoE | Microsoft's Agentic Center of Enablement, the preview-stage product packaging agent governance into a Power Platform solution. | Citizen-dev Copilot agents still need pro-dev ALM |
| ALM | Application Lifecycle Management, the discipline of solutions, environments, and pipelines that promotes a build from Dev to Prod. | Agent Skills are the cheapest ALM your team will ever buy, Citizen-dev Copilot agents still need pro-dev ALM, Governance is the new agent bottleneck, SharePoint is the grounded RAG layer you already own, Models depreciate, eval suites compound |
| ANN | Approximate Nearest Neighbor search, the algorithm class behind vector retrieval, including HNSW and DiskANN. | Your agent's identity is a Postgres role |
| Azure Durable Task | Azure Functions runtime feature that journals every workflow step for replay and exactly-once execution. | Pick the durable runtime before your second agent ships, Agent Flows are the SMB onramp to agentic automation, Copilot Studio Workflows is the spine LLM agents needed |
| Bedrock AgentCore | AWS Bedrock managed agent runtime with sandboxed identity, an agent registry, and CloudTrail audit trails. | Agent Flows are the SMB onramp to agentic automation, Four agent patterns recur regardless of vendor stack, Copilot Studio Workflows is the spine LLM agents needed, Stop deploying agents, start onboarding them |
| BM25 | Classical lexical ranking function used alongside dense vectors in hybrid search to recover keyword precision. | Dataverse is the agent backend most M365 shops already own, Structured retrieval beats vector RAG for enterprise agents, Your agent's identity is a Postgres role |
| BYOUI | Bring Your Own UI, embedding a Copilot Studio agent through a custom web client via the M365 Agents SDK. | Copilot Studio: most agent problems are integration problems |
| canary | Rollout slice that receives a new variant first while scorers compare it against baseline before promoting. | Production agents need runtime scorers, not just pre-ship evals |
| Chroma | Open-source embedding database used in the Context Rot study and as a local-first vector store. | Memory tiers, not bigger models, cut your agent token bill |
| chunking | Splitting source documents into retrievable units sized for embedding, ranking, and the model context window. | SharePoint is the grounded RAG layer you already own, Structured retrieval beats vector RAG for enterprise agents |
| citizen developer | Business user who builds apps or agents on low-code platforms without writing production code. | Agent Flows are the SMB onramp to agentic automation, Governance is the new agent bottleneck |
| Claude | Anthropic's Claude model family (Haiku, Sonnet, Opus), the Anthropic-hosted LLM default for production agents. | Mittelstand AI runs on a Mac Studio, not someone else's cloud, Agent Skills are the cheapest ALM your team will ever buy, Pick the durable runtime before your second agent ships, Dataverse is the agent backend most M365 shops already own, Production agents need runtime scorers, not just pre-ship evals, Stop buying Copilot seats, rebuild one process instead, Stop routing every agent step through a frontier LLM, Memory tiers, not bigger models, cut your agent token bill, Four agent patterns recur regardless of vendor stack, Structured retrieval beats vector RAG for enterprise agents, Microsoft 365 ships agent inventory not observability, Temperature zero will not save you, Models depreciate, eval suites compound |
| Cloud Flows | Power Automate workflow that runs in the cloud without LLM nodes, the deterministic sibling of Agent Flows. | Agent Flows are the SMB onramp to agentic automation, Copilot Studio Workflows is the spine LLM agents needed |
| CoE | Center of Excellence or Enablement, the governance hub for Power Platform makers and agents inside a tenant. | Citizen-dev Copilot agents still need pro-dev ALM |
| Conditional Access | Entra policy engine that gates sign-in by device compliance, network location, and risk signal per identity. | Governance is the new agent bottleneck, Stop deploying agents, start onboarding them, Copilot Studio: most agent problems are integration problems, Microsoft 365 ships agent inventory not observability |
| Connection References | Power Platform indirection layer letting a Solution carry connector bindings across environments without re-authoring. | Citizen-dev Copilot agents still need pro-dev ALM |
| context window | Maximum input tokens a model can attend to in one request, the upper bound on stuffable history plus retrieved context. | Mittelstand AI runs on a Mac Studio, not someone else's cloud, Memory tiers, not bigger models, cut your agent token bill, Structured retrieval beats vector RAG for enterprise agents |
| Copilot Control System | Microsoft 365 Admin Center surface giving three tenant levers over agent access, sharing, and publishing. | Governance is the new agent bottleneck |
| Copilot Credits | Tenant-pooled metered unit Microsoft bills for Copilot Studio agent interactions, tool calls, and retrieval. | Citizen-dev Copilot agents still need pro-dev ALM, Agent Flows are the SMB onramp to agentic automation, Copilot Studio Workflows is the spine LLM agents needed, Copilot Studio: most agent problems are integration problems, Microsoft 365 ships agent inventory not observability |
| Copilot Studio | Microsoft low-code builder for conversational agents and Agent Flows on top of Power Platform and M365. | Pick the durable runtime before your second agent ships, Boring ML on Fabric still beats LLM-everything for the Mittelstand, Dataverse is the agent backend most M365 shops already own, Production RL is finally cheap enough to close the agent loop, Stop buying Copilot seats, rebuild one process instead, Stop routing every agent step through a frontier LLM, Citizen-dev Copilot agents still need pro-dev ALM, Agent Flows are the SMB onramp to agentic automation, Governance is the new agent bottleneck, Four agent patterns recur regardless of vendor stack, Copilot Studio Workflows is the spine LLM agents needed, SharePoint is the grounded RAG layer you already own, Copilot Studio: most agent problems are integration problems, Structured retrieval beats vector RAG for enterprise agents, Microsoft 365 ships agent inventory not observability, Temperature zero will not save you, Models depreciate, eval suites compound |
| critique shadowing | Eval-judge calibration pattern where an expert authors pass/fail critiques on 100+ examples until human-judge agreement passes 90%. | – |
| Dataverse | Microsoft typed, governed table store with row-level security that underpins Power Platform and Copilot Studio agents. | Boring ML on Fabric still beats LLM-everything for the Mittelstand, Dataverse is the agent backend most M365 shops already own, Stop buying Copilot seats, rebuild one process instead, Citizen-dev Copilot agents still need pro-dev ALM, Agent Flows are the SMB onramp to agentic automation, Governance is the new agent bottleneck, Four agent patterns recur regardless of vendor stack, SharePoint is the grounded RAG layer you already own, Structured retrieval beats vector RAG for enterprise agents |
| DBOS | Python library that turns the Postgres a team already operates into a durable execution runtime via decorators. | Pick the durable runtime before your second agent ships |
| Defender | Microsoft Defender, the threat-detection surface that feeds the MAC agent risk taxonomy. | Stop deploying agents, start onboarding them, Microsoft 365 ships agent inventory not observability |
| deterministic spine | Code or graph that owns control flow and treats every LLM call as a replayable, checkpointed activity. | Pick the durable runtime before your second agent ships, Production RL is finally cheap enough to close the agent loop, Stop buying Copilot seats, rebuild one process instead, Stop routing every agent step through a frontier LLM, Agent Flows are the SMB onramp to agentic automation, Four agent patterns recur regardless of vendor stack, Copilot Studio Workflows is the spine LLM agents needed |
| Direct Line | Bot Framework REST channel used to embed Copilot Studio agents in custom web clients or server-side callers. | Copilot Studio: most agent problems are integration problems |
| DiskANN | Approximate nearest neighbor index that keeps the graph on NVMe with a small RAM cache, used by Azure pg_diskann. | Your agent's identity is a Postgres role |
| DLP | Data Loss Prevention policy that blocks sensitive data egress or restricts which connectors a flow may invoke. | Dataverse is the agent backend most M365 shops already own, Citizen-dev Copilot agents still need pro-dev ALM, Agent Flows are the SMB onramp to agentic automation, Governance is the new agent bottleneck, Stop deploying agents, start onboarding them |
| DPO | Direct Preference Optimization, a fine-tuning algorithm that trains on preference pairs without a separate reward model. | Production RL is finally cheap enough to close the agent loop |
| drift | Silent shift in user inputs, retrieval corpus, or vendor model behavior that pulls production away from the test snapshot. | Agent Skills are the cheapest ALM your team will ever buy, Production agents need runtime scorers, not just pre-ship evals, Stop routing every agent step through a frontier LLM, Copilot Studio Workflows is the spine LLM agents needed, Structured retrieval beats vector RAG for enterprise agents, Temperature zero will not save you, Models depreciate, eval suites compound |
| DSPM | Data Security Posture Management, the Purview surface that surfaces oversharing and AI exposure across a tenant. | SharePoint is the grounded RAG layer you already own |
| durable execution | Runtime pattern that journals every workflow step so the system replays from the last checkpoint after a crash. | Pick the durable runtime before your second agent ships, Production RL is finally cheap enough to close the agent loop, Four agent patterns recur regardless of vendor stack, Copilot Studio Workflows is the spine LLM agents needed |
| embeddings | Dense vector representations of text or other content used by retrieval and semantic search systems. | Memory tiers, not bigger models, cut your agent token bill, Four agent patterns recur regardless of vendor stack, SharePoint is the grounded RAG layer you already own, Structured retrieval beats vector RAG for enterprise agents, Your agent's identity is a Postgres role, Models depreciate, eval suites compound |
| Entra Agent ID | Microsoft directory identity for an AI agent, with blueprint, principal, and instance tiers governing lifecycle. | Citizen-dev Copilot agents still need pro-dev ALM, Governance is the new agent bottleneck, Four agent patterns recur regardless of vendor stack, Stop deploying agents, start onboarding them, Microsoft 365 ships agent inventory not observability |
| Entra ID | Microsoft Entra ID, the Azure identity directory (formerly Azure AD) that issues principals for users, apps, and agents. | Dataverse is the agent backend most M365 shops already own, Stop routing every agent step through a frontier LLM, Citizen-dev Copilot agents still need pro-dev ALM, Governance is the new agent bottleneck, Four agent patterns recur regardless of vendor stack, Stop deploying agents, start onboarding them, Copilot Studio: most agent problems are integration problems, Structured retrieval beats vector RAG for enterprise agents, Microsoft 365 ships agent inventory not observability, Your agent's identity is a Postgres role |
| Environment Variables | Power Platform Solution primitive that stores per-environment configuration values (URLs, IDs) for promotion across dev, test, prod. | Mittelstand AI runs on a Mac Studio, not someone else's cloud, Citizen-dev Copilot agents still need pro-dev ALM, Observability outlives your agent framework |
| EU AI Act | EU regulation classifying AI systems by risk tier, with GPAI obligations enforced from 2026-08-02. | Boring ML on Fabric still beats LLM-everything for the Mittelstand, Governance is the new agent bottleneck, Microsoft 365 ships agent inventory not observability |
| eval suite | Versioned bank of test cases and scorers that grades an agent across intent, grounding, action, and conversation layers. | Production RL is finally cheap enough to close the agent loop, Production agents need runtime scorers, not just pre-ship evals, Stop routing every agent step through a frontier LLM, Memory tiers, not bigger models, cut your agent token bill, Observability outlives your agent framework, Temperature zero will not save you, Models depreciate, eval suites compound |
| Fabric Data Agent | Microsoft Fabric NL-to-SQL agent that translates plain-language questions into structured queries against OneLake sources. | Boring ML on Fabric still beats LLM-everything for the Mittelstand |
| fine-tuning | Adapting a base model's weights on task-specific data, typically via SFT, DPO, or reinforcement learning. | Production RL is finally cheap enough to close the agent loop, Temperature zero will not save you, Models depreciate, eval suites compound |
| Foundry | Microsoft Azure AI platform for building, hosting, observing, and tuning agents across frameworks. | Pick the durable runtime before your second agent ships, Boring ML on Fabric still beats LLM-everything for the Mittelstand, Production RL is finally cheap enough to close the agent loop, Production agents need runtime scorers, not just pre-ship evals, Stop routing every agent step through a frontier LLM, Governance is the new agent bottleneck, Observability outlives your agent framework, Your agent's identity is a Postgres role |
| Frontier Tuning | Microsoft's managed RL fine-tuning surface that runs inside the customer's Azure tenant boundary. | Production RL is finally cheap enough to close the agent loop |
| function calling | Model output convention that emits structured JSON arguments for a named tool the orchestrator then executes. | – |
| GDPR | EU General Data Protection Regulation; Art. 32 governs technical and organisational measures for any LLM processing personal data. | Mittelstand AI runs on a Mac Studio, not someone else's cloud |
| golden set | Held-out labeled examples used as a regression bench that catches model and prompt regressions before release. | Production agents need runtime scorers, not just pre-ship evals |
| GPAI | General-Purpose AI under the EU AI Act, foundation models subject to systemic-risk obligations from 2026-08-02. | Governance is the new agent bottleneck |
| GPT | OpenAI's generative pre-trained transformer family; GPT-4, GPT-5, and o-series reasoning models are the production cloud baseline. | Mittelstand AI runs on a Mac Studio, not someone else's cloud, Boring ML on Fabric still beats LLM-everything for the Mittelstand, Production RL is finally cheap enough to close the agent loop, Stop buying Copilot seats, rebuild one process instead, Stop routing every agent step through a frontier LLM, Memory tiers, not bigger models, cut your agent token bill, SharePoint is the grounded RAG layer you already own, Structured retrieval beats vector RAG for enterprise agents, Microsoft 365 ships agent inventory not observability, Temperature zero will not save you, Your agent's identity is a Postgres role |
| grounding | Constraining a model's answer to cited content retrieved from a known corpus or typed data source. | Dataverse is the agent backend most M365 shops already own, Agent Flows are the SMB onramp to agentic automation, Governance is the new agent bottleneck, Four agent patterns recur regardless of vendor stack, Copilot Studio Workflows is the spine LLM agents needed, SharePoint is the grounded RAG layer you already own, Copilot Studio: most agent problems are integration problems, Microsoft 365 ships agent inventory not observability, Temperature zero will not save you, Models depreciate, eval suites compound |
| GRPO | Group Relative Policy Optimization, a critic-free RL algorithm popularized by DeepSeek-R1 that halves training compute. | Production RL is finally cheap enough to close the agent loop, Models depreciate, eval suites compound |
| guardrails | Deterministic checks around model input and output that block disallowed content or unsafe tool calls. | Four agent patterns recur regardless of vendor stack, Observability outlives your agent framework, Copilot Studio: most agent problems are integration problems, Microsoft 365 ships agent inventory not observability |
| hallucination | Plausible but wrong model output not supported by the retrieved sources or the underlying ground truth. | Stop buying Copilot seats, rebuild one process instead |
| HNSW | Hierarchical Navigable Small World, an in-memory ANN graph index used by pgvector and most vector databases. | Dataverse is the agent backend most M365 shops already own, Your agent's identity is a Postgres role |
| human-in-the-loop | Workflow pattern that pauses an agent at a designated approval moment before an irreversible action. | Four agent patterns recur regardless of vendor stack, Copilot Studio Workflows is the spine LLM agents needed, Stop deploying agents, start onboarding them |
| hybrid search | Retrieval that combines BM25 lexical and dense-vector ranking, often with a reranker on top. | Memory tiers, not bigger models, cut your agent token bill, Your agent's identity is a Postgres role |
| Inspect AI | Open-source agent eval framework from the UK AI Safety Institute with sandboxed execution and multi-turn loops. | Production agents need runtime scorers, not just pre-ship evals, Models depreciate, eval suites compound |
| KV cache | Cached key-value tensors from prior tokens so each new generation step does not recompute attention over the entire prefix. | – |
| LangChain | Python/JS framework for composing LLM calls, retrievers, and tools; parent project of LangGraph and LangSmith. | Pick the durable runtime before your second agent ships, Memory tiers, not bigger models, cut your agent token bill, Observability outlives your agent framework, Copilot Studio Workflows is the spine LLM agents needed, Temperature zero will not save you, Your agent's identity is a Postgres role, Models depreciate, eval suites compound |
| Langfuse | Open-source LLM observability platform that ingests OTLP traces and runs model-based runtime evals. | Production RL is finally cheap enough to close the agent loop, Production agents need runtime scorers, not just pre-ship evals, Observability outlives your agent framework |
| LangGraph | LangChain's durable state-graph runtime for agents with checkpointed nodes, edges, and shared state. | Pick the durable runtime before your second agent ships, Stop routing every agent step through a frontier LLM, Memory tiers, not bigger models, cut your agent token bill, Agent Flows are the SMB onramp to agentic automation, Four agent patterns recur regardless of vendor stack, Observability outlives your agent framework, Copilot Studio Workflows is the spine LLM agents needed, Models depreciate, eval suites compound |
| LangSmith | LangChain's hosted trace and eval platform, native to LangGraph and LangChain agents. | Production RL is finally cheap enough to close the agent loop, Production agents need runtime scorers, not just pre-ship evals, Observability outlives your agent framework, Temperature zero will not save you |
| LLM | Large Language Model, a transformer-based generative model that emits text and tool calls from prompts. | Mittelstand AI runs on a Mac Studio, not someone else's cloud, Pick the durable runtime before your second agent ships, Boring ML on Fabric still beats LLM-everything for the Mittelstand, Production RL is finally cheap enough to close the agent loop, Production agents need runtime scorers, not just pre-ship evals, Stop buying Copilot seats, rebuild one process instead, Stop routing every agent step through a frontier LLM, Memory tiers, not bigger models, cut your agent token bill, Agent Flows are the SMB onramp to agentic automation, Four agent patterns recur regardless of vendor stack, Observability outlives your agent framework, Copilot Studio Workflows is the spine LLM agents needed, Stop deploying agents, start onboarding them, Copilot Studio: most agent problems are integration problems, Structured retrieval beats vector RAG for enterprise agents, Microsoft 365 ships agent inventory not observability, Temperature zero will not save you, Your agent's identity is a Postgres role, Models depreciate, eval suites compound |
| LLM-as-judge | Pattern that uses a model to score another model's output against a rubric, calibrated against human labels. | Production agents need runtime scorers, not just pre-ship evals, Temperature zero will not save you, Models depreciate, eval suites compound |
| M365 | Microsoft 365, the SaaS bundle including Outlook, Teams, SharePoint, OneDrive, and the Copilot surfaces. | Mittelstand AI runs on a Mac Studio, not someone else's cloud, Boring ML on Fabric still beats LLM-everything for the Mittelstand, Dataverse is the agent backend most M365 shops already own, Production RL is finally cheap enough to close the agent loop, Stop buying Copilot seats, rebuild one process instead, Stop routing every agent step through a frontier LLM, Citizen-dev Copilot agents still need pro-dev ALM, Agent Flows are the SMB onramp to agentic automation, Governance is the new agent bottleneck, Four agent patterns recur regardless of vendor stack, Observability outlives your agent framework, Copilot Studio Workflows is the spine LLM agents needed, SharePoint is the grounded RAG layer you already own, Stop deploying agents, start onboarding them, Copilot Studio: most agent problems are integration problems, Structured retrieval beats vector RAG for enterprise agents, Microsoft 365 ships agent inventory not observability, Your agent's identity is a Postgres role |
| M365 Agents SDK | Microsoft SDK in Python, JS, and C# for embedding Copilot Studio agents with Entra SSO and Tenant Graph Grounding. | Copilot Studio: most agent problems are integration problems |
| Managed Environments | Power Platform tier that enforces premium licensing, tiered DLP, weekly digests, and admin governance on a tenant environment. | Citizen-dev Copilot agents still need pro-dev ALM, Agent Flows are the SMB onramp to agentic automation, Governance is the new agent bottleneck |
| MCP | Model Context Protocol, a standardized tool-surface specification agents use to call databases, files, and services. | Agent Skills are the cheapest ALM your team will ever buy, Boring ML on Fabric still beats LLM-everything for the Mittelstand, Dataverse is the agent backend most M365 shops already own, Stop buying Copilot seats, rebuild one process instead, Citizen-dev Copilot agents still need pro-dev ALM, Agent Flows are the SMB onramp to agentic automation, Four agent patterns recur regardless of vendor stack, Observability outlives your agent framework, SharePoint is the grounded RAG layer you already own, Stop deploying agents, start onboarding them, Structured retrieval beats vector RAG for enterprise agents, Your agent's identity is a Postgres role |
| memory tiers | Working, episodic, semantic, and archival memory layers wrapped around stateless LLM calls to bound token cost. | Memory tiers, not bigger models, cut your agent token bill |
| Mittelstand | German mid-market firms (typically 20 to 500 employees) that anchor the SMB AI procurement and architecture argument. | Mittelstand AI runs on a Mac Studio, not someone else's cloud, Agent Skills are the cheapest ALM your team will ever buy, Pick the durable runtime before your second agent ships, Boring ML on Fabric still beats LLM-everything for the Mittelstand, Dataverse is the agent backend most M365 shops already own, Production RL is finally cheap enough to close the agent loop, Production agents need runtime scorers, not just pre-ship evals, Stop buying Copilot seats, rebuild one process instead, Stop routing every agent step through a frontier LLM, Citizen-dev Copilot agents still need pro-dev ALM, Memory tiers, not bigger models, cut your agent token bill, Agent Flows are the SMB onramp to agentic automation, Governance is the new agent bottleneck, Four agent patterns recur regardless of vendor stack, Observability outlives your agent framework |
| MLflow | Open-source experiment tracking and model registry library, shipped natively inside Microsoft Fabric workspaces. | Boring ML on Fabric still beats LLM-everything for the Mittelstand |
| MoE | Mixture-of-Experts architecture where a router activates a small subset of expert weights per token to lower inference cost. | Mittelstand AI runs on a Mac Studio, not someone else's cloud |
| n8n | Self-hostable workflow automation tool that wires LLM nodes alongside conventional triggers and connectors. | Stop buying Copilot seats, rebuild one process instead, Agent Flows are the SMB onramp to agentic automation, Four agent patterns recur regardless of vendor stack, Models depreciate, eval suites compound |
| Nemotron | NVIDIA's family of small specialist models trained from scratch for tool use and agentic reasoning. | Stop routing every agent step through a frontier LLM |
| NL2SQL | Natural language to SQL, the agent capability that turns plain-English questions into structured queries against a typed schema. | Boring ML on Fabric still beats LLM-everything for the Mittelstand |
| OBO | On-Behalf-Of OAuth flow that lets middleware call a downstream API as the originally signed-in user instead of the service principal. | Copilot Studio: most agent problems are integration problems |
| OData | Open Data Protocol query syntax Dataverse uses to express typed filters against governed tables. | Dataverse is the agent backend most M365 shops already own, Structured retrieval beats vector RAG for enterprise agents |
| Ollama | Local inference runtime that serves open-weight models behind a CLI compatible with Copilot, Codex, and IDE tools. | Mittelstand AI runs on a Mac Studio, not someone else's cloud |
| OneLake | Microsoft Fabric's single storage substrate where lakehouses and warehouses share one tenant-scoped namespace. | Boring ML on Fabric still beats LLM-everything for the Mittelstand |
| OpenInference | Arize-authored agent trace schema, the OpenTelemetry-compatible alternative to the OTel GenAI semconv. | Observability outlives your agent framework |
| OpenTelemetry | Open observability standard whose GenAI semantic conventions define spans for LLM, tool, and agent steps. | Pick the durable runtime before your second agent ships, Production agents need runtime scorers, not just pre-ship evals, Observability outlives your agent framework |
| PAYG | Pay-As-You-Go metered billing for Copilot Studio consumption above prepaid Copilot Credit packs. | Boring ML on Fabric still beats LLM-everything for the Mittelstand, Stop buying Copilot seats, rebuild one process instead, Citizen-dev Copilot agents still need pro-dev ALM, Microsoft 365 ships agent inventory not observability |
| pgvector | Postgres extension that adds vector columns, HNSW and IVFFlat indexes, and hybrid search for embeddings. | Dataverse is the agent backend most M365 shops already own, Memory tiers, not bigger models, cut your agent token bill, Your agent's identity is a Postgres role |
| Phoenix | Arize-authored MIT-licensed OpenTelemetry-native LLM observability and eval framework. | Production RL is finally cheap enough to close the agent loop, Production agents need runtime scorers, not just pre-ship evals, Observability outlives your agent framework |
| PII | Personally Identifiable Information, the data class agent guardrails and DLP policies are primarily configured to redact or block. | Production RL is finally cheap enough to close the agent loop, Governance is the new agent bottleneck, Copilot Studio: most agent problems are integration problems, Microsoft 365 ships agent inventory not observability |
| Pinecone | Managed vector database with project-scoped API keys, namespace isolation, and a serverless tier for multi-tenant agents. | Dataverse is the agent backend most M365 shops already own, SharePoint is the grounded RAG layer you already own, Structured retrieval beats vector RAG for enterprise agents, Your agent's identity is a Postgres role |
| Power Platform | Microsoft low-code suite (Power Apps, Power Automate, Power BI, Copilot Studio, Dataverse) for business automation. | Dataverse is the agent backend most M365 shops already own, Stop buying Copilot seats, rebuild one process instead, Citizen-dev Copilot agents still need pro-dev ALM, Agent Flows are the SMB onramp to agentic automation, Governance is the new agent bottleneck, Four agent patterns recur regardless of vendor stack, Copilot Studio Workflows is the spine LLM agents needed, SharePoint is the grounded RAG layer you already own, Structured retrieval beats vector RAG for enterprise agents, Microsoft 365 ships agent inventory not observability, Temperature zero will not save you, Models depreciate, eval suites compound |
| Power Platform Pipelines | Power Platform in-product promotion path between linked environments, replacing the deprecated ALM Accelerator. | Stop routing every agent step through a frontier LLM, Citizen-dev Copilot agents still need pro-dev ALM, Structured retrieval beats vector RAG for enterprise agents, Models depreciate, eval suites compound |
| Power Platform Solution | Dataverse-tracked package of Power Platform components (agents, connectors, variables) that promotes between environments. | – |
| PPO | Proximal Policy Optimization, the on-policy RL algorithm with a critic network that GRPO replaced for many LLM fine-tunes. | Production RL is finally cheap enough to close the agent loop |
| pro-code | Engineering path that ships code, tests, and infrastructure as files in git rather than configuration in a designer. | Citizen-dev Copilot agents still need pro-dev ALM |
| prompt caching | Provider feature that stores prefix tokens so repeated system prompts and RAG context bill at a fraction of the input rate. | Memory tiers, not bigger models, cut your agent token bill, Structured retrieval beats vector RAG for enterprise agents |
| prompt injection | Adversarial input crafted to override an agent's system prompt or tool-use policy. | Microsoft 365 ships agent inventory not observability |
| Purview | Microsoft compliance suite covering audit, eDiscovery, DLP, and AI data security across the tenant. | Citizen-dev Copilot agents still need pro-dev ALM, Governance is the new agent bottleneck, SharePoint is the grounded RAG layer you already own, Stop deploying agents, start onboarding them, Structured retrieval beats vector RAG for enterprise agents, Microsoft 365 ships agent inventory not observability |
| Qdrant | Open-source vector database with first-class multi-tenancy through named collections and quota controls. | Dataverse is the agent backend most M365 shops already own, SharePoint is the grounded RAG layer you already own, Your agent's identity is a Postgres role |
| RAG | Retrieval-Augmented Generation, the pattern of fetching external context and stuffing it into the prompt at query time. | Mittelstand AI runs on a Mac Studio, not someone else's cloud, Dataverse is the agent backend most M365 shops already own, Production agents need runtime scorers, not just pre-ship evals, Memory tiers, not bigger models, cut your agent token bill, Four agent patterns recur regardless of vendor stack, SharePoint is the grounded RAG layer you already own, Structured retrieval beats vector RAG for enterprise agents |
| Rego | Open Policy Agent declarative language used by ACS and other engines to express runtime authorization rules. | Observability outlives your agent framework |
| reranker | Second-pass ranking model that reorders top-K retrieval candidates before they enter the LLM prompt. | SharePoint is the grounded RAG layer you already own, Structured retrieval beats vector RAG for enterprise agents |
| Restate | Self-hostable durable execution runtime that shapes workflow steps as journaled HTTP-style RPC calls. | Pick the durable runtime before your second agent ships, Production RL is finally cheap enough to close the agent loop |
| reward hacking | Failure mode where an RL-trained agent maximises the grader's score while violating the intended objective. | Production RL is finally cheap enough to close the agent loop, Models depreciate, eval suites compound |
| RFT | Reinforcement Fine-Tuning, OpenAI's managed RL training pipeline priced at roughly one hundred dollars per training hour. | Production RL is finally cheap enough to close the agent loop |
| RL | Reinforcement Learning, training that updates a model from a scalar reward signal scored on trajectories or outcomes. | Production RL is finally cheap enough to close the agent loop, Production agents need runtime scorers, not just pre-ship evals, Stop routing every agent step through a frontier LLM |
| RLAIF | Reinforcement Learning from AI Feedback, using an AI judge instead of human labels as the reward signal. | Production RL is finally cheap enough to close the agent loop |
| RLHF | Reinforcement Learning from Human Feedback, training a reward model from human preference labels and then policy-optimising against it. | – |
| RLS | Row-Level Security, database-enforced predicate that restricts which rows a session can read or write. | Dataverse is the agent backend most M365 shops already own, Structured retrieval beats vector RAG for enterprise agents, Your agent's identity is a Postgres role |
| SFT | Supervised Fine-Tuning on curated input-output pairs, the cheapest path to specialize a base model. | Production RL is finally cheap enough to close the agent loop |
| shadow AI | Unsanctioned AI agents, browser tools, or API keys that staff wire up outside the governance surface. | Agent Flows are the SMB onramp to agentic automation, Governance is the new agent bottleneck, Microsoft 365 ships agent inventory not observability |
| SharePoint | Microsoft document store and collaboration surface whose tenant semantic index grounds Copilot agents. | Dataverse is the agent backend most M365 shops already own, Stop buying Copilot seats, rebuild one process instead, Citizen-dev Copilot agents still need pro-dev ALM, Agent Flows are the SMB onramp to agentic automation, Governance is the new agent bottleneck, Four agent patterns recur regardless of vendor stack, SharePoint is the grounded RAG layer you already own, Copilot Studio: most agent problems are integration problems, Microsoft 365 ships agent inventory not observability, Temperature zero will not save you, Models depreciate, eval suites compound |
| SLM | Small Language Model, typically 4B to 8B parameters, purpose-trained for tool use and structured output. | – |
| Step Functions | AWS state-machine orchestrator used as a deterministic spine around Bedrock agents. | Pick the durable runtime before your second agent ships, Agent Flows are the SMB onramp to agentic automation, Four agent patterns recur regardless of vendor stack, Copilot Studio Workflows is the spine LLM agents needed |
| structured output | Model output constrained to a typed schema via tool calling or JSON mode for reliable downstream parsing. | Stop routing every agent step through a frontier LLM |
| system prompt | Top-of-prompt instructions that set role, tone, tool policy, and refusal behavior for every turn. | Stop routing every agent step through a frontier LLM, Memory tiers, not bigger models, cut your agent token bill, Observability outlives your agent framework, Stop deploying agents, start onboarding them, Copilot Studio: most agent problems are integration problems |
| Temporal | Managed durable execution platform whose workflows journal every activity for replay and audit. | Pick the durable runtime before your second agent ships, Production RL is finally cheap enough to close the agent loop, Stop buying Copilot seats, rebuild one process instead, Agent Flows are the SMB onramp to agentic automation, Four agent patterns recur regardless of vendor stack, Copilot Studio Workflows is the spine LLM agents needed |
| Tenant Graph Grounding | Copilot Studio feature that grounds answers in the signed-in user's SharePoint and Microsoft Graph data, exposed via the M365 Agents SDK. | Governance is the new agent bottleneck, Four agent patterns recur regardless of vendor stack, Copilot Studio: most agent problems are integration problems |
| tool use | Model capability to emit named function calls with typed arguments that the orchestrator executes and feeds back. | Agent Skills are the cheapest ALM your team will ever buy, Stop routing every agent step through a frontier LLM, Four agent patterns recur regardless of vendor stack |
| top-K | Retrieval parameter selecting the K nearest neighbours from a vector index before reranking or generation. | Structured retrieval beats vector RAG for enterprise agents |
| vector search | Nearest-neighbor retrieval over embedding vectors using cosine or dot-product similarity. | Memory tiers, not bigger models, cut your agent token bill, Structured retrieval beats vector RAG for enterprise agents, Your agent's identity is a Postgres role |
| Weaviate | Open-source vector database whose multiTenancyConfig provides tenant-level HNSW shard isolation. | Dataverse is the agent backend most M365 shops already own |
| Work IQ | Microsoft semantic layer where emails, Teams messages, SharePoint files, and Dataverse rows are one addressable surface. | Agent Skills are the cheapest ALM your team will ever buy, Stop buying Copilot seats, rebuild one process instead, SharePoint is the grounded RAG layer you already own |