We've now run the second AI wave session in Sydney, Hobart, Melbourne and online, and a lot of new vocabulary goes past in an hour. One piece of feedback from the online session asked for a glossary, which is a fair request. This is it.
Most of these terms sit on Power Platform and Copilot Studio. The ones under "The action" are where our Experimentation pillar starts: getting people trying agents for real, inside limits, before anyone commits to a big build.
The trends
Personal productivity agents
Chat-style assistants that help one person do their own work faster: answering policy questions, drafting, summarising. Microsoft 365 Copilot is the usual example, and agents like these are by far the most common kind today. They're the front door to AI for most organisations.
They speed up one person's step. The process around that person still moves at the pace of its slowest handoff.
Process agents
Agents applied inside a business process, often spanning several systems and data sources: reading a contract that lands in a shared mailbox, checking it against a register, drafting the reply. They're spreading across organisations (brand voice, email coaching) and down into departments such as finance, marketing and legal. Microsoft calls this the operational long tail.
This is the second wave. The real gain comes when the whole matter is automated end to end.
Agent and agency
An agent is software given instructions, tools and the agency to decide how to use those tools to reach a goal, inside the guardrails it has been set. A script follows a fixed path; an agent chooses its path.
Agentic automation
Automating a business process with a mix of agents, for the steps that need judgement, and ordinary automation, for the steps that don't. It's process re-engineering with AI in the places that need it.
Copilot Credits
The unit Microsoft uses to meter Copilot Studio agents and workflows. Different actions consume different numbers of credits - a generated answer, a tool call, a workflow step - so what an agent costs to run comes down to how it's built. Credits can be allocated to an environment and capped per agent.
The cost of AI is now visible per run, which is a good thing. It also means a badly designed agent shows up on the bill.
- Microsoft Learn: Billing rates and management
- Microsoft Learn: Overview of usage-based billing - on the GitHub Copilot harness, building consumes credits too
- Ours: The three ways to buy Copilot Credits and Your agent has a price tag now
Tokens
The pieces of text a language model reads and writes, roughly three quarters of a word each. Models are priced, limited and measured in tokens.
- Microsoft Learn: Understand tokens
Context window
The most text a model can take in at once: the instructions, the conversation so far, documents and tool results. The largest models now take around a million tokens, which is several long novels.
Turns
An agent working on a task doesn't answer in one go. It plans, acts, checks the result and decides what to do next, calling the model again each time. Each of those cycles is a turn, and a single run can take many.
An agent's instructions are sent with every turn. Six pages of instructions across thirty turns is a lot of text, and a lot of credits.
Instructions
The always-on guidance an agent is given: its role, its tone, its limits. One of Microsoft's engineers put it bluntly at the Copilot Studio architecture bootcamp in London: instructions are abused by everybody, and developers are the worst. The temptation is to write every rule and every if-statement into them.
Instructions are a soft guardrail. They shape behaviour without enforcing it, and because they ride along on every turn, long instructions cost both money and accuracy. Keep them to what is always true, and move the rest into skills, tools or workflow steps.
Context rot
The drop in quality when a model is given too much to hold at once. Quality tends to fall well before the context window is full: the model starts missing things, drifting off track and making errors.
Bloated instructions and a stuffed context window give you higher cost and worse answers at the same time.
Build is no longer the constraint
Building an app or an agent is faster than it has ever been. The constraint now is the people who know the process - the steps, the systems, the data and the business rules in their heads - and getting their time.
The technology
Model
The large language model doing the reasoning: a GPT model, Claude, or another. In the session's analogy, the model is the engine.
Harness
The runtime around the model. It feeds the model its inputs, controls the outputs, gives it tools and decides how it plans. If the model is the engine, the harness is the car built around it - chassis, steering wheel and body. Copilot Studio offers three harnesses, and the choice shapes how an agent behaves, how it works with people and how it's billed.
- Microsoft Learn: Harnesses in Copilot Studio
- Microsoft Learn: Compare Copilot Studio agent harnesses
Copilot chat harness and Agent Builder
The harness behind Microsoft 365 Copilot. When you create an agent with Agent Builder inside Copilot, you're building on it. It suits agents grounded in one user's own world - their calendar, their documents, Microsoft Graph - such as employee Q&A. Agent Builder and Copilot Studio are two doors into the same platform.
- Microsoft Learn: Agent Builder overview
Work IQ Dev Tools
A pro-code toolset, in preview, for developers building declarative agents on the Copilot chat harness from the command line and VS Code.
- Microsoft Learn: Compare tools for declarative agents
Microsoft Graph
The single API over Microsoft 365 data: mail, calendars, files, people and Teams. Agents that need a user's own work context reach it through Graph.
- Microsoft Learn: Overview of Microsoft Graph
Standard harness and topics
What Copilot Studio was until recently, and the evolution of Power Virtual Agents. It works through topics - conversation paths triggered by what the user asks - and plans once, then executes. That makes it fast and predictable for turn-by-turn conversations such as a helpdesk password reset, a customer service enquiry or Q&A at scale. Microsoft has said it isn't going away.
- Microsoft Learn: Create and edit topics
GitHub Copilot harness
The newest harness in Copilot Studio. Despite the name, it's separate from the GitHub Copilot coding product. It works in a loop - look at the latest state, decide, act, observe the result, critique it, go again - much like the thinking you see in deep research tools. That makes it better at recovering from errors and adapting when conditions change. Skills and memory are only available here.
- Microsoft Learn: Agents powered by the GitHub Copilot harness
Deterministic and non-deterministic steps
A deterministic step gives the same output for the same input every time: if the purchase order is over $10,000, route it to the finance manager. A non-deterministic, or agentic, step needs judgement and could reasonably come out differently - working out who is really behind a trust that signed a contract, for example.
Deterministic steps are faster, cheaper and more predictable. Most processes are mostly deterministic, with a handful of steps that genuinely need reasoning.
The right level of intelligence at every step
The guiding principle from the bootcamp: use the right level of intelligence for each step, no less and no more. Rules handle the rules, an agent handles the judgement, and each AI step gets the model that suits it.
Workflows
A new kind of automation in Copilot Studio, built to work alongside agents. Think of it as Power Automate evolved for agents. A workflow carries the deterministic backbone of a process - lookups, calculations, routing - and calls AI only at the steps that need it. Existing Power Automate cloud flows keep running; workflows are for new builds.
- Microsoft Learn: Workflows overview
Workflow nodes
The AI building blocks inside a workflow. An agent node calls an agent and takes back a structured result. A classify node sorts an input into categories you define with examples. An extract node reads a document and pulls out the fields you choose. A Copilot node calls Microsoft 365 Copilot or one of its agents. Each AI node can use its own model.
- Microsoft Learn: Agent node · Classify node · Extract node · Microsoft 365 Copilot node (preview)
Human review and human in the loop
Human in the loop means a person stays part of the process at the points that need them: sign-offs, ambiguous decisions, and anything costly or hard to undo. In a workflow, the human review step pauses the process and asks a person for a decision or information by email or in Teams, with buttons to respond, then carries on with their answer.
An instruction that says "always ask first" can be ignored. A human review step can't be skipped, which makes it a hard guardrail.
- Microsoft Learn: Request information from human review in workflows
Skills
A packaged set of instructions, and optionally templates and scripts, for doing one task the same way every time - drafting a deficiency letter, say. Skills standardise how a task is done, and they keep the agent lean, because the full skill is only read when it's needed. Skills follow an open format, so a skill written for one tool can often be used in another.
- Microsoft Learn: Skills overview for agents
Front matter
The name and short description at the top of a skill. Only the front matter is sent with every turn, so the agent always knows the skill exists and when to use it without carrying the whole thing around.
Progressive disclosure
Giving an agent a summary up front and the detail only when it asks for it. Skills work this way, and it's the main defence against both rising credit costs and context rot.
Hooks
Points in an agent's life - a prompt arriving, a tool about to run, an error - where a workflow can run every time, and in some cases block the action. We mentioned them as coming soon in the session; they're now in preview.
- Microsoft Learn: Hooks (preview)
Connected agents
One agent handing a self-contained task to another agent and getting a result back. They're useful for large areas of work that need their own tools and skills. Splitting small tasks across many agents doesn't make them more accurate.
- Microsoft Learn: Add other agents overview
Memory
An agent remembering a user's preferences and past interactions, so a conversation picked up days later carries on where it left off. It's per user and per agent, stored in your tenant, and deleted after a period of inactivity. Process status belongs in a database such as Dataverse.
- Microsoft Learn: Memory (preview)
Choosing the model per step
Copilot Studio lets you pick which model an agent or an AI step uses, including models from OpenAI and Anthropic. A small, fast model can classify an email, and a stronger one can handle the step that needs real reasoning. Keeping the model separate from the workflow also means you can swap in better models as they arrive without rebuilding the process. How you work stays yours.
- Microsoft Learn: Select a primary AI model for your agent
Anthropic models and data handling
A question from the session: is it safe to use different frontier models with sensitive information? In Microsoft 365 and Copilot Studio, Anthropic operates as a Microsoft subprocessor, and administrators can turn it off at tenant level. The organisations we see doing that usually cite data sovereignty more than data security.
- Microsoft Learn: Anthropic models in Microsoft Online Services
Shared mailbox trigger
Many processes start with an email landing in a shared mailbox. A Power Automate cloud flow can trigger on that email, save the attachments and hand them to a workflow.
- Microsoft Learn: Overview of cloud flows · Trigger a cloud flow based on email properties
Dataverse
The database underneath Power Platform and Dynamics 365, with tables, security and business logic in one governed place. Because agents, workflows, apps and portals all read and write the same Dataverse data, they fit together without integration work.
- Microsoft Learn: What is Microsoft Dataverse?
Supervision app
Our term for the app people use to oversee automated work: what came in, what the agent did, which items are waiting on a person, and where each person fits in the overall process. Not every process needs one. For some, a review request in email or Teams is enough.
Code apps
A newer style of Power Apps: apps written in code, such as React, but hosted, secured and governed in Power Platform, with access to Dataverse and connectors. You get full control over how the app looks - the demo showed the same screen as a timeline, a pipeline, a terminal and a neon disco.
- Microsoft Learn: Power Apps code apps overview
Solution aware, and dev, test, prod
A solution-aware component can be packaged into a Power Platform solution and moved between environments, typically development, test and production. That's how code apps, agents and workflows are maintained with ordinary software lifecycle practice, including source control.
- Microsoft Learn: Application lifecycle management overview
Power Pages and client portals
Power Pages builds external-facing websites on Dataverse. Because the automated process already keeps its data there, adding a signed-in portal for clients, contractors or brokers is a short step.
- Microsoft Learn: What is Power Pages? · Authentication in Power Pages
The action
Governed experimentation
Letting people try agents for real, inside limits: a dedicated environment, only the connectors and data you allow, and a capped credit budget. People get room to learn while the risk and the spend stay bounded.
- Microsoft Learn: Environments overview · Managed environments overview · Configure data policies for agents
Capped credit budget
Allocating a fixed number of Copilot Credits to an environment, and setting limits per agent, so experimentation can't run up an open-ended bill.
- Microsoft Learn: Manage Copilot Credits and capacity
AI FinOps
Applying financial operations discipline - visibility, allocation, forecasting and optimisation - to AI spend. It's an emerging practice. In short: know what each agent costs per run, decide what a run is worth, and design to that.
- Microsoft Learn: What is FinOps? · Manage AI costs
Cost certainty
Knowing what a process will cost to run, within a narrow range. An agent doing everything can vary widely from run to run. Moving the deterministic steps into a workflow brings both the cost and the variation down.
Shared mailbox audit
The quickest way to find automation candidates. Look at what arrives in your shared mailboxes, how much of it there is, and what people do with each item. A lot of work hides behind them.
Value driver, cost ceiling and quality bar
Three questions to ask the person who owns a process before automating it. First, what actually matters to them - effort hours, elapsed time, or risk? Often it's elapsed time. Second, what would they pay per run to have it done? Third, how often is the manual work reworked today? If one in five needs rework, the quality bar is 80%, and anything better than that is a gain.
Governed proof of concept
A short, time-boxed test of the specific technical things that must be true for a project to go ahead: a few weeks, a defined list of validations, and a clear go or no-go at the end. Increasingly, we fold the proof into the first weeks of the build instead.
Three amigos
A small decision group that meets often during a build: a delegate from the client who can make decisions, a product owner, and a developer or solution architect. Small questions get answered in the room within days, without waiting for a meeting of ten or fifteen people.
Vibe coding
Building software by describing what you want to an AI tool and refining the result through conversation. It has made prototypes cheap and fast. Production still needs ordinary discipline: review, source control, and dev, test and prod.
Citizen development
People outside IT - the ones who know the work - building their own apps, agents and automations with low-code and AI tools, while IT helps govern, harden and support what they build.
- Microsoft Learn: Power Platform adoption guidance · Fostering fusion team culture
Forward-deployed engineering
Placing a skilled engineer who can talk both business and technology inside a business unit, measured on an outcome such as reducing partner admin time. Because they're measured on the outcome, they're trusted to build and deploy, within agreed governance. It's the successor to embedding business analysts who brought ideas back to a central backlog.
Build the machine that builds the machine
Another phrase from the bootcamp. When you build an agent, also build the reusable tooling - skills for gathering requirements, for building, for testing - that makes the next build faster and better. Our own version is a suite of build skills that improve as we use them.
Terms and features in this area change quickly. These definitions are current as at October 2026, and the Microsoft Learn pages are the authority on what is generally available and what is still in preview.