Copilot Studio Agents: Quick Start Guide for M365

What's new — June to September 2026

Agents read scanned PDFs (GA)
Newly created declarative agents ground answers in
scanned PDFs and image-based documents referenced
from SharePoint (Roadmap 550514, late August 2026).

Agent sharing amongst makers
Makers will share agents built in the GitHub Copilot
harness, with Agent Viewer or Editor access and
auto-assigned roles (Roadmap 569475, in development).

Connectors configured in chat
Makers will configure an agent’s connectors through
an in-chat sign-in instead of a full settings page
(Roadmap 569930, in development).

Copilot Cowork is GA
Generally available worldwide as of June 16, 2026
(was Frontier). Needs an M365 Copilot licence;
usage billed separately.

Work IQ APIs reached GA
June 16, 2026 — the context layer your Studio
agents use to ground in M365 data. This is what
triggers usage metering for custom agents.

Copilot Credits usage billing
New unified consumption meter. Custom Studio
agents grounding in M365 data are metered at
$0.01/credit (pay-as-you-go). See Usage billing below.

Agent-node in workflows (Preview)
Embed agent reasoning directly as a workflow step
via the new agent node in Copilot Studio, reducing
manual handoffs (Roadmap 566998).

SharePoint lists as knowledge (Preview)
Ground agents natively in SharePoint list data
(tasks, inventory, customers) without moving the
data (Roadmap 566859).

Two credential-safety previews
Block maker-provided credentials for AI agents
(Roadmap 566997) and detect credential
oversharing at publish/share time (Roadmap 566873).

Roadmap consolidation
From July 2, 2026, Copilot Studio and agent
roadmap updates publish to the Microsoft 365
Roadmap; Release Planner is no longer the source.

Scheduled prompts for agents (GA)
July 2026 release notes: recurring prompts can
now target declarative agents such as Analyst
and Idea Coach, replacing manual initiation for
ongoing tasks or insights.

Agent Store submissions (GA)
As of July 15, 2026, makers can submit Agent
Builder agents to the Agent Store's "Built by
your org" section, gated by admin review and
approval in the Microsoft 365 admin center.

Frontier agent approval flow (GA)
Frontier and Microsoft-built agents now appear
in the Copilot Control System agent inventory;
blocked users can request access and admins
approve or reject (Roadmap 494809, July 10, 2026).

Express mode for flows (Preview)
Accelerates flow runs invoked by agents and
apps, reducing timeout failures; the roadmap
lists GA for November 2026 (Roadmap 566999,
status as of July 10, 2026).

Run-only agent sharing (Preview)
Makers can share autonomous agents with end
users on a run-only basis, without granting edit
permissions or ownership (Roadmap 567894).

Agent 365 agents in Chat (In development)
Agents published to Agent 365 are managed in the
Microsoft 365 admin center; those implementing the
Agent2Agent protocol can act as connected agents
in Agent Builder or Copilot Studio (Roadmap 567670).

Declarative agent manifest 1.8
Schema version 1.8 adds an EmailActions capability
for email write operations such as triage, supervised
send, delete, inbox rules, auto-reply, and folder
management, separate from the read-only Email capability.

Harnesses documented (GA)
Copilot Studio now documents three harnesses, the
runtime layer between the agent you design and the
model that reasons: the GitHub Copilot harness, the
standard harness, and the Copilot chat harness.

GitHub harness billing (GA)
Agents and workflows on the GitHub Copilot harness
bill by usage in Copilot Credits: LLM tokens, tools
including knowledge and MCP, and the harness itself.
Three charge lines, not one.

Credit enforcement policy (GA)
With no allocated credits, or all of them consumed,
any experience that needs credits stops until capacity
is added. End users see agents stop responding, so set
alerts below the limit rather than at it.

Agent Readiness (In development)
A single always-visible Review indicator in the build
experience flags what is blocking or at risk, from
policy restrictions to missing evaluations, before it
derails a publish (Roadmap 568762).

Work IQ MCP servers (Preview)
A Microsoft 365 Copilot licence is required to use them,
and admins allow or block MCP servers across the
organisation from the Microsoft 365 admin center. No GA
date is stated and no roadmap ID is cited.

Dataverse as a knowledge source (In development)
Dataverse becomes a native knowledge source in Copilot
Studio, so agents can ground in structured enterprise
data: records, customer data, operational data and
business application content (Roadmap 568929).

Azure SQL knowledge source (In development)
Adds Azure SQL to the enterprise data sources available
when building agents in Copilot Studio (Roadmap 568930).

Agents and skills from the plus menu (In development)
Users insert agents and skills into a prompt from an
Agents and Skills tab, reached with Add work content on
the plus menu, or by typing / or @ in the prompt itself
(Roadmap 569213).

GitHub Copilot harness (GA)
The harness, Copilot Studio’s new foundation for
building reasoning-heavy agents and workflows, reached
general availability on 2 September 2026. Workflows,
MCP servers as agent tools, and the Windows 365 for
Agents MCP server are GA with it.

Skills and primary models (GA)
From the same release: Skills package modular, reusable
instructions across agents, and Claude Sonnet 5 and
GPT-5.5 Chat are generally available as primary models.

Claude Fable 5.1 (Rolling out)
Anthropic’s Claude Fable 5.1 joins the frontier models
in Copilot Cowork and Copilot Studio. Rollout to eligible
users began 1 September 2026; availability varies by
region and organisation, and admins manage access
through the Microsoft 365 admin center.

Where Fable 5.1 fits
Microsoft positions the model for long-running work,
financial analysis, and front-end visual coding, planning
more tightly up front and summarising more concisely at
the end, for longer tasks delegated in Cowork and Studio.
Work IQ grounds responses in files, meetings, chats, and
business data within existing permissions.

MCP Apps in agents (In development)
Agents will render interactive app experiences inline via
MCP Apps, an extension of the Model Context Protocol:
a UI-enabled tool from a compatible MCP server can show
sortable tables and drilldowns in the conversation, with
a plain-text fallback elsewhere (Roadmap 570433).
Preview target October CY2026; GA target November CY2026.

Human approval for tool calls (In development)
A per-tool, per-agent toggle lets makers require human
approval before an agent runs specific tools: a gated call
pauses with a request describing the intended action, and
a person approves, approves for the session, or denies.
Requests appear inline in Teams and Microsoft 365 Copilot
(Roadmap 570434). Target GA per the roadmap:
September CY2026.

App-first agent creation (In development)
Makers will start with the application their agent works
with and configure the capabilities it needs, bringing
tools and knowledge into one place (Roadmap 569474).
Target GA per the roadmap: October CY2026.

Chat-native agent responses (In development)
Agent responses in Copilot Chat will look, feel, and
behave like Copilot Chat’s own: similar formatting,
compressed chain of thought, smoother streaming, resume
and retry, voice dictation, clickable citations, and tool
execution traces (Roadmap 570435). Target GA per the
roadmap: September CY2026.

GPT-6 Astra (Rolling out)
OpenAI’s GPT-6 Astra joins the model choices in
Copilot Cowork and Copilot Studio, rolling out from
4 September 2026; availability varies by region and
organisation. Work IQ grounds it in files, meetings, chats,
and business data within existing permissions. Admins
control access in the Microsoft 365 admin center, so the
model policy belongs in the rollout plan (Microsoft Copilot
blog, 4 September 2026).

Build apps in Cowork and Studio (Preview)
Cowork’s /app skill builds business apps from natural
language and is available through the Frontier programme;
App (Preview) in Copilot Studio does the same and rolls out
in public preview over the week following 10 September 2026.
Apps use connectors and Work IQ, follow Entra and connector
policies, and show in the admin center inventory once
published. Building and running them follows the
usage-based billing model.

Custom skills in declarative agents (Preview)
Declarative agents can carry packaged skills: a folder with
a SKILL.md instruction file, added in Agent Builder (up to
eight per agent, 50 MB per package) or the Agents Toolkit.
Preview for Frontier Preview organisations; scripts run in a
sandbox, sensitivity labels are honoured, and admin
governance is available. Decide who may upload skill
packages (Microsoft Learn, published 3 September 2026).

Cost visibility in Copilot Studio (In development)
Makers will see consumption costs in three places: per
evaluation, broken down across evaluation generation,
test runs and model grading (Roadmap 571195); per agent
in the Monitor tab, with an estimated breakdown across
the agent lifecycle (Roadmap 571196); and for preview and
testing usage in the preview tab and agent history,
covering every historical run (Roadmap 571194). The
roadmap lists all three as in development for
worldwide multi-tenant.

Line-of-business MCP apps (Learn guidance)
New guidance for developers building MCP apps that
extend declarative agents. Inside a declarative agent, an
MCP app adds an interactive app widget to an MCP tool
response. Three downstream authentication options are
set out: delegated identity, application identity and
shared credential (Microsoft Learn, 18 September 2026).

Claude Opus 5.5 and GPT-6 Sol (rolling out)
Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Sol are
rolling out to Microsoft Copilot as of 22 September 2026,
across Word, Excel, PowerPoint, Chat, Cowork and Copilot
Studio. Microsoft presents the change as expanding model
choice in Copilot (Microsoft Copilot blog).

At a glance

What it is
A named Copilot instance with a fixed role and
optional knowledge sources. Always available via @.

What makes it different
Prompts require someone to remember to run them.
Agents are always there — consistent, team-wide.

Effort to deploy
Paste an instruction block into Copilot Studio.
Name it. Save it. Done. ~15 minutes first time.

Agent type	What it does	Best for
Role-specific assistant	Acts as an expert in a domain (project control, procurement, HSE) with the right vocabulary and boundaries	Teams with a recurring domain question type
Knowledge base agent	Answers questions from pinned documents — procedures, standards, policies	Replacing "where do I find the procedure for X?"
Document producer	Generates structured outputs (reports, checklists, briefs) from a consistent template	Recurring deliverables with a fixed format
Review agent	Reviews documents against criteria — checks for gaps, flags issues, scores against a rubric	Pre-submission review, proposal quality checks

Agent vs prompt — when to use each

One-off · Personal

Prompt in Chat

Used once or occasionally
Context changes every time (different files, different people)
Personal task — only you need the result
Exploring and iterating — output format still evolving
Requires specific M365 data (email, calendar, meetings)

Recurring · Team-wide

Deployed Agent

Used regularly by multiple people
Same role and instructions every time
Team needs consistent output format and boundaries
Output should not depend on who runs it
Grounded on shared documents (procedures, standards)

Deploying an agent in 5 steps — no coding required

Get the instruction block

Go to github.com/kesslernity/awesome-copilot-studio-agents. Find the agent for your use case. Copy the full content of the ## Instructions section.

Open the agent builder

Go to m365.cloud.microsoft/chat/agent/new. This is the Copilot Studio agent builder — no admin portal required for basic deployments.

Paste and name the agent

Paste the instruction block into the Instructions field. Give the agent a name (max 30 characters). Add conversation starters from the repo's ## Conversation Starters section.

Add knowledge sources (optional)

Under Knowledge, add SharePoint URLs or upload files. The agent will answer from those sources in addition to its instructions. Leave blank for instruction-only agents.

Save and share

Click Save. The agent appears in your Copilot Chat. To share with the team: Share → Copy link or publish to a Teams channel. Licensed users can @mention it immediately.

Recommended agents to deploy first

From github.com/kesslernity/awesome-copilot-studio-agents · 73 agents · 13 domains

Project Status Reporter

Project Manager · Delivery Lead · PMO

Project Management

What it does

Generates RAG status reports from project data
Flags schedule, cost, and risk deviations
Formats output for executive review or team distribution
Maintains consistent structure across all projects

How to use it

ProjectStatusReporter Generate the weekly status report for the Coral Bay project."
ProjectStatusReporter Flag every item that moved from Amber to Red this week."
ProjectStatusReporter Draft the ExCom summary paragraph for this project."

Ground this agent on your project SharePoint folder for context-aware outputs. Add the project brief and risk register as knowledge sources.

Pipeline Review Preparer

Sales Manager · Business Development · Commercial Lead

Sales & BD

What it does

Converts CRM export data into structured pipeline review
Applies RAG health status and coverage ratios
Flags deals at risk by stage age and last activity
Produces format ready for a leadership meeting

How to use it

PipelineReviewPreparer [paste CRM export] — generate the pipeline review for this week's call."
PipelineReviewPreparer Which deals have had no activity in 14 days?"
PipelineReviewPreparer What is our pipeline coverage ratio against Q2 target?"

Paste CRM export data directly into the chat — this agent does not need CRM integration. Outputs structured review, not probability predictions.

Meeting Brief Generator

Manager · Executive · EA · Any role with recurring external meetings

Productivity

What it does

Generates a pre-meeting brief in a consistent format
Pulls context from emails, meetings, and pinned documents
Surfaces open commitments and likely discussion topics
Outputs under 200 words — designed for 5-minute reading

How to use it

MeetingBriefGenerator I have a call with [name] at [company] in 30 minutes. Generate my brief."
MeetingBriefGenerator What open items do I have with this contact from our last 3 exchanges?"

Best results when you reference the contact via @name and the meeting via @meeting name in the same prompt.

Procedure Reference Assistant

Any team with shared procedures, standards, or compliance documents

Knowledge Base

What it does

Answers questions from a fixed set of pinned documents
Points to specific sections and page numbers
Replaces "who do I ask about X?" for known procedures
Consistent answers team-wide — no individual interpretation

How to use it

ProcedureRef What is the approval threshold for a change order under our project execution procedure?"
ProcedureRef Which document covers subcontractor qualification requirements?"
ProcedureRef List the HSE notification steps for a first-aid incident."

Add your procedure documents, standards, and policy files as knowledge sources when deploying. No instruction block needed beyond the role definition — the knowledge sources do the work.

Rules to follow

1Instruction blocks must stay under 8,000 characters. Copilot Studio enforces this limit. Blocks from the repo are pre-checked, but if you edit them, keep an eye on length.
2Name the agent clearly — it becomes the @mention. Max 30 characters. Choose a name the team will remember and type correctly. Avoid generic names like "Assistant".
3Ground on documents, not on the agent's memory. Agents do not retain conversation history across sessions. For knowledge that should persist, add it as a knowledge source (SharePoint URL or uploaded file).
4Test before sharing. Run 5–10 prompts covering the intended use cases and at least 2 edge cases before distributing the @mention link to the team.
5Agents should state what they will not do. The instruction block should include explicit boundaries (e.g., "Do not approve documents. Do not provide legal advice."). This prevents the agent from being used outside its intended scope.

Usage billing — what a deployed agent costs (new, June 2026)

As of the June 16, 2026 GA wave, a Copilot Studio agent that grounds in your M365 data via the Work IQ APIs (GA June 16, 2026) is metered in Copilot Credits — Microsoft's unified consumption currency. The Copilot licence is the access gate; agent runtime is billed on top. A task's cost is the sum of model use, context retrieval, tool calls, and runtime. For a full reference on how Copilot Studio agents are billed, see the cost model.

Copilot Credits (pay-as-you-go)

$0.01 per credit. Commit-for-discount pricing is also available ("P3").

What a task costs

Model use + context retrieval + tool calls + runtime — summed per task.

Off by default

Usage billing is off until an admin enables it. Admins set spend limits and usage alerts at tenant, group, or user level.

Reporting

A new Cost Management dashboard in the M365 admin center reports Credits consumption.

Models

Claude Opus 5 in the preview orchestrator (since 24 July 2026), plus Claude Opus 4.8 and Claude Sonnet 5. Anthropic models require admin opt-in in the Microsoft 365 admin center. Check the live list on the agent's Overview page under Model before costing a build.

Agent 365 ($15/user/mo)

Covers governance and identity only — each agent gets an Entra Agent ID. Agent runtime is billed separately via Copilot Credits / Foundry tokens. The $15 seat does not include runtime.

Not separately metered

Pre-built Copilot agents

Microsoft's own pre-built Copilot agents under an existing licence
Covered by the $30/user/month Copilot add-on (unchanged)
No Copilot Credits meter applies to these

Metered in Copilot Credits

Custom & third-party agents

Custom Copilot Studio agents that ground in M365 data via Work IQ
Microsoft Foundry and raw Work IQ API agents
A Copilot licence is still required; the Credits meter applies on top of it

Grace period: tenants that had a Frontier user between March 30 and June 16, 2026 are not billed for Cowork usage until July 1, 2026. Source: Microsoft Learn — Copilot Credits usage-based billing.

Limits and constraints

Declarative agents in Copilot Studio are grounded on instructions and knowledge sources only. They do not run code, call external APIs, or perform actions without additional configuration.

Instruction block limit

8,000 characters. Enforced by Copilot Studio.

Knowledge sources

SharePoint URLs and uploaded files. No external web URLs.

Conversation memory

No cross-session memory. Each conversation starts fresh.

Sharing

Share via link or Teams channel. Recipient needs M365 Copilot licence to use.

Permissions

Agent respects each user's access rights. It cannot surface content a user cannot access directly.

Updating an agent

Edit the instruction block and save. All @mention users get the updated agent immediately — no redeployment or re-sharing required.

Actions (e.g. send email, create task)

Not available in basic declarative agents. Requires Power Automate integration — contact your Microsoft admin or check your organisation's licence portal.