🏗️ Agent Builder (Copilot Studio Lite)¶
Last updated: 2026-04-08 | Status: 🟢 Active — features evolving rapidly
What is Agent Builder?¶
☕ Think of it like this...
Imagine you're a café owner (M365 Copilot is your café). Agent Builder is like a self-service kiosk builder — it lets you create specialised kiosks (agents) that help customers with specific tasks. You don't need to be an engineer to build one. You just describe what the kiosk should do, tell it where to find information, and place it in your café.
Copilot Studio is like hiring a professional kiosk manufacturer — they can build anything, connect to external systems, handle complex workflows, and deploy kiosks to multiple locations. More power, more complexity.
Agent Builder is the no-code agent creation experience built directly into Microsoft 365 Copilot. It creates what Microsoft calls Declarative Agents — custom AI assistants that run on the same platform, models, and security as M365 Copilot itself.
Key Facts¶
| Aspect | Detail |
|---|---|
| Official name | Agent Builder in Microsoft 365 Copilot |
| Also known as | Copilot Studio Lite, CS Lite |
| What it creates | Declarative Agents |
| Cost | Included with M365 Copilot license (no extra cost) |
| Storage | No Dataverse storage consumed |
| Coding required | None — fully no-code |
| Available at | microsoft365.com/chat, office.com/chat, Teams Desktop/Web |
| NOT available | Mobile apps, Teams Chat (known limitation) |
Where Does It Sit?¶
┌─────────────────────────────────────────────────────────┐
│ Microsoft 365 Copilot │
│ ┌───────────────────┐ ┌─────────────────────────────┐ │
│ │ Agent Builder │ │ Copilot Studio (Full) │ │
│ │ (CS Lite) │ │ │ │
│ │ │ │ │ │
│ │ • No-code │ │ • Low-code / Pro-code │ │
│ │ • Quick & simple │──│ • Advanced workflows │ │
│ │ • M365 only │ │ • Multi-channel deploy │ │
│ │ • Knowledge focus │ │ • API integrations │ │
│ │ │ │ • Custom actions │ │
│ └───────────────────┘ └─────────────────────────────┘ │
│ ▲ ▲ │
│ │ "Copy to Studio" │ │
│ └──────────────────────────┘ │
│ Start simple → Promote when needed │
└─────────────────────────────────────────────────────────┘
✅ All Capabilities (April 2026)¶
🗂️ A. Knowledge Sources¶
These are the data sources your agent can use to answer questions. Think of them as the reference books your agent can read.
| Source | Limit | Details |
|---|---|---|
| SharePoint Sites | Up to 20 sites | 200 files, 50 folders, 10 subfolder levels per source |
| OneDrive Files | Part of SP limit | Files and folders from OneDrive |
| File Uploads | Up to 20 files | Stored in SharePoint Embedded containers. Max 512MB (DOCX/PPT/PDF) or 150MB (TXT/CSV/XLS) |
| Web URLs | Up to 4 URLs | Public websites only, 2 levels deep, no query parameters |
| Graph Connectors | 50M items included | ServiceNow, Salesforce, SAP, websites, databases, LOB systems |
| Teams Chat | Up to 5 threads | Specific chat threads as knowledge |
| All accessible | ⚠️ Cannot scope — agent sees all permitted email | |
| File Groups | 🔵 Preview | Group multiple files as a single logical source |
| Metadata-filtered SharePoint | 🔵 Preview | Filter SP content by metadata for better relevance |
| Tenant Graph Grounding | 🔵 Preview | Semantic search across tenant's Microsoft Graph data |
Key Gotcha: Email Scoping
Unlike SharePoint (where you pick specific sites), email cannot be scoped. Your agent sees everything the user has access to. Use detailed instructions to tell the agent which emails to focus on (e.g., "Only reference emails from the last 7 days about Project Alpha").
📝 B. Instructions & Configuration¶
This is the brain of your agent — how you tell it what to do and how to behave.
| Feature | What It Does |
|---|---|
| Natural Language Instructions | Describe the agent's role, tone, guardrails, what to do and NOT do. 8,000 character limit. |
| Starter/Suggested Prompts | Pre-built prompts users can click to get started. Up to 12 prompts per agent (as of April 2026 in Corp). |
| Custom Icon & Name | Brand your agent with a name and avatar |
| Describe Tab | Natural language agent definition (AI helps configure it) |
| Configure Tab | Manual control over knowledge sources, prompts, settings |
💡 Pro Tip: Instructions Matter A LOT
The quality of your agent is 80% determined by the instructions you write. Be specific about:
- Who the agent is (persona/role)
- What it should and shouldn't do
- How it should respond (tone, format, length)
- What knowledge to prioritise
- When to say "I don't know" vs. making assumptions
Why Instructions Are the #1 Thing to Get Right¶
☕ Think of it like this...
Instructions are like the training manual you give to a new café employee on their first day. If you say "just be helpful", they'll make it up as they go. If you give them a clear manual — "greet customers by name, never recommend items we're out of, always ask about allergies, speak casually but professionally" — they'll be brilliant from day one. Same with agents.
The instructions field is where you define your agent's entire personality, scope, behaviour, and guardrails. Everything the agent does flows from this. A vague instruction = a vague agent.
Instructions Hard Limit: 8,000 Characters¶
| Aspect | Detail |
|---|---|
| Character limit | 8,000 characters maximum (including spaces) |
| Format | Plain text — but you CAN use Markdown headings (##), bullet points (-), and numbered lists (1.) for structure |
| Language | Write in the language you want the agent to respond in (English for NZ use cases) |
| Updates | You can edit instructions anytime after creation — changes take effect immediately |
⚠️ Gotcha: 8,000 characters fills up fast
With detailed personas, rules, topics, and examples — 8,000 characters is roughly 1,200-1,400 words. Plan your instructions carefully. If you need more space, focus on the most critical rules and let the knowledge sources handle the detail.
Best Practices for Writing Instructions¶
| Practice | Why It Matters | Example |
|---|---|---|
| Start with a role statement | Sets the agent's identity immediately | "You are the NZ Policy Advisor — a helpful assistant that..." |
| Use clear section headings | Helps the LLM parse your intent (it reads structure, not just words) | ## Your Role, ## Rules You MUST Follow, ## Topics You Cover |
| Be explicit about what NOT to do | Prevents scope creep and hallucinations | "Do NOT answer immigration questions", "Do NOT make up answers" |
| Add a mandatory disclaimer | Critical for legal/compliance agents | "ALWAYS include: This is general information only, not legal advice" |
| Specify response format | Controls output quality | "Use bullet points", "Keep answers under 200 words", "Always cite the source" |
| Handle uncertainty explicitly | Prevents the agent from guessing | "If you're not sure, say so clearly — never fabricate information" |
| Define tone | Consistency across all responses | "Be conversational but professional", "Avoid legal jargon" |
| Include examples (if space allows) | Shows the agent exactly what good looks like | "Example: When asked about sick leave, respond with the entitlement, eligibility, and a link" |
Instructions Structure Template¶
Here's a reusable template structure that works well for most agents:
You are [Agent Name] — [one-sentence description of who you are and what you do].
## Your Role
- [Primary responsibility]
- [How you communicate]
- [Tone and style]
## Rules You MUST Follow
1. [Most important rule — scope boundary]
2. [Citation/sourcing rule]
3. [Uncertainty handling rule]
4. [Disclaimer requirement]
5. [Format/style rules]
## Topics You Cover
- [Topic area 1]
- [Topic area 2]
- [Topic area 3]
## Topics You Do NOT Cover
- [Out-of-scope topic 1]
- [Out-of-scope topic 2]
## Response Format
- [How to structure answers]
- [Length expectations]
- [Citation format]
Common Instruction Mistakes¶
| ❌ Mistake | Why It's Bad | ✅ Better Approach |
|---|---|---|
| "Be helpful" | Too vague — every agent is "helpful" | "Help NZ business owners understand employment law by citing official sources" |
| No scope boundaries | Agent will try to answer EVERYTHING | "ONLY answer questions about NZ employment, business, tax, or H&S" |
| No uncertainty handling | Agent will hallucinate when it doesn't know | "If unsure, say 'I don't have that information' — never guess" |
| Writing a novel | 8,000 chars fills fast; LLM may lose focus on key rules | Prioritise rules over examples; keep it tight |
| All caps everything | Dilutes emphasis | Use ALL CAPS only for critical rules: "ALWAYS", "NEVER", "MUST" |
| Contradictory rules | Confuses the LLM | Review instructions for conflicts before saving |
| Soft boundary language | LLM overrides it to be "helpful" | Use CAPS + exact fallback text (see technique below) |
🔑 Boundary Enforcement Technique (Learned from A03 Build)¶
⚠️ Real-World Learning
During the A03 (Brand Voice Coach) build, we discovered that soft boundary rules get ignored by the LLM. The agent's natural helpfulness overrides gentle instructions.
The problem: We wrote "Don't write content from scratch without a draft" — but when tested, the agent happily wrote a marketing email from scratch anyway. It wanted to be helpful.
The fix: Three techniques combined make boundaries stick:
1. ALL CAPS for critical keywords
The LLM treats capitalised keywords as higher priority. Use sparingly for maximum effect.
Best CAPS keywords: ALWAYS, NEVER, ONLY, MUST, DO NOT
2. Provide exact fallback text
Don't just say "decline" — tell the agent exactly what to say when it hits a boundary. This removes the temptation to improvise a helpful answer.
❌ Weak: "Don't write content from scratch"
✅ Strong: "NEVER write content from scratch. If asked, respond with:
'I'm your Brand Voice Coach — I transform existing text
into our brand voice. Please give me a draft to work with,
and I'll make it shine! ✨'"
3. Positive redirect
End the fallback with what the user SHOULD do instead. This feels helpful (satisfying the LLM's instinct) while still enforcing the boundary.
❌ Just decline: "I can't do that"
✅ Redirect: "I can't write from scratch, but give me any draft
and I'll transform it into our brand voice!"
💡 The Three-Part Formula
For every boundary rule, use this pattern:
Example:
Test your boundaries! After building any agent, always test the "Topics You Do NOT Cover" with dedicated boundary prompts (see agent A01 and A03 for examples). If a boundary fails, strengthen it with this technique.
🎨 C. Skills (Office Artifact Creation)¶
These are the superpowers your agent has beyond just answering questions.
| Skill | What It Does | Status |
|---|---|---|
| Document Generation | Creates Word docs, Excel sheets, PowerPoint decks directly in chat | ✅ GA (Dec 2025) |
| Code Interpreter | Runs Python code for data analysis, charts, calculations | ✅ GA |
| Image Generation | Creates images from text descriptions | ✅ GA |
| Web Search | Searches the public web for real-time information | ✅ GA (admin-controlled) |
The master toggle for artifacts is: "Generate documents, charts, and code" in Agent Builder settings.
User: "Create a weekly status report for Project Alpha"
│
▼
┌──────────────────────┐
│ Agent processes │
│ request using: │
│ │
│ 📂 SharePoint data │
│ 📝 Instructions │
│ 🎨 Skills │
│ │
└──────────┬───────────┘
│
┌──────────┴───────────┐
│ Generates: │
│ │
│ 📄 Word report │
│ 📊 Excel data │
│ 📽️ PPT slides │
└──────────────────────┘
🔧 D. Tools & Actions¶
Tools let your agent do things beyond just reading and answering — it can take actions in M365 apps.
| Feature | Status | What It Does |
|---|---|---|
| Tool Groups | ✅ GA | Add Outlook, SharePoint, and other M365 tools in one step |
| Input Widgets | 🔵 Preview | Calendar picker, file picker, dropdowns in chat |
| Auto SSO Detection | ✅ GA | Automatic sign-on for connected services |
| Entra ID Consent | 🔵 Preview | Authentication consent prompts inside chat |
| Human-in-the-Loop | 🔵 Preview | Pause agent → collect approval via Outlook → resume workflow |
| Copilot Actions | ✅ GA | Scheduled/automated "set and forget" prompts (e.g., daily summaries) |
What are Tool Groups?
Tool Groups are pre-packaged sets of actions for M365 apps. For example, the Outlook Tool Group gives your agent the ability to:
- Draft emails
- Search inbox
- Create calendar events
- Send meeting invites
The SharePoint Tool Group lets your agent:
- Search documents
- Read file contents
- Navigate sites and libraries
🤝 E. Sharing & Deployment¶
| Target | Supported | Notes |
|---|---|---|
| Just me (private) | ✅ | Default — only you can use it |
| Specific people | ✅ | Share with named colleagues |
| Security groups | ✅ | Share with AD/Entra security groups |
| Everyone in org | ✅ | Org-wide deployment |
| M365 Copilot Chat | ✅ | Primary interface |
| Teams Desktop/Web | ✅ | Via @mention in Copilot |
| Teams Chat | ❌ | Known limitation |
| Mobile apps | ❌ | Not available yet |
| External channels (web, Slack) | ❌ | Use Copilot Studio |
🔒 F. Governance & Security¶
| Feature | Detail |
|---|---|
| Permission inheritance | Agent respects SharePoint, OneDrive, and email permissions — users only see what they have access to |
| Sensitivity labels | Honoured from source documents |
| DLP policies | M365 DLP applies to agent responses |
| Admin controls | IT can enable/disable Agent Builder tenant-wide |
| Entra Agent Identity | 🔵 Preview — automatic agent-level identity for auditing |
| Dataverse storage | Not consumed (no extra cost) |
| Feedback system | Built-in thumbs up/down + detailed feedback submission |
🚀 G. Advanced Features¶
| Feature | Status | What It Does |
|---|---|---|
| Multi-Agent Orchestration | 🔵 Preview | Chain multiple specialised agents together for complex tasks |
| Copy to Copilot Studio | ✅ GA | Promote your agent to Studio for advanced capabilities |
| Agent Store | ✅ GA | Discover and deploy pre-built agents from Microsoft + partners |
| Template Hub | 🔵 Preview | Start from pre-built templates for common scenarios |
| Agent Launchpad | 🔵 Preview | Centralised agent discovery for end users |
⚠️ Limitations & Workarounds¶
Quick Reference Table¶
| # | Limitation | Impact | Workaround |
|---|---|---|---|
| 1 | Max 20 SharePoint sites per agent | Can't cover massive document estates | Split into multiple agents by domain; aggregate docs into fewer sites |
| 2 | Max 4 web URLs, 2 levels deep | Limited public web coverage | Extract web content to PDF/DOCX and upload as files instead |
| 3 | Max 20 uploaded files per agent | Can't upload large document sets | Consolidate into fewer, larger documents; use SharePoint instead |
| 4 | Email can't be scoped | Agent may reference irrelevant emails | Use instructions to specify email focus (e.g., "only emails about Project X from last 7 days") |
| 5 | No multi-step workflows | Can't build complex business processes | Promote to Copilot Studio; or use Copilot Actions for simple scheduling |
| 6 | No custom API calls | Can't connect to non-Graph systems natively | Use Graph connectors for supported systems; promote to Studio for custom APIs |
| 7 | No topic-tree design | Can't build guided multi-turn conversations | Write very detailed instructions to simulate guided flows |
| 8 | Not available on mobile | Mobile users can't access custom agents | Use Teams desktop/web or M365 Copilot web app |
| 9 | No Teams Chat support | Can't @mention agent in Teams conversations | Use agent in M365 Copilot Chat interface instead |
| 10 | No Lockbox or CMK | May not meet highest compliance requirements | Evaluate customer risk tolerance; consider Studio for regulated industries |
| 11 | SP auto-share only with security groups | Can't auto-share with "everyone" | Manually set permissions on SP files/folders for all-org access |
| 12 | No custom UI widgets | Can't create rich interactive cards | Promote to Studio for Adaptive Cards and custom UIs |
The Promotion Path (Key Customer Story)¶
┌─────────────────────┐ ┌──────────────────────┐ ┌────────────────────────┐
│ START HERE │ │ GROW HERE │ │ SCALE HERE │
│ │ │ │ │ │
│ Agent Builder │────▶│ Copy to Copilot │────▶│ Full Copilot Studio │
│ (CS Lite) │ │ Studio │ │ │
│ │ │ │ │ │
│ • No-code │ │ • Add workflows │ │ • API integrations │
│ • Quick prototype │ │ • Add actions │ │ • Multi-channel │
│ • Knowledge focus │ │ • Add connectors │ │ • Enterprise scale │
│ • Minutes to build │ │ • Hours to enhance │ │ • Days to weeks │
└─────────────────────┘ └──────────────────────┘ └────────────────────────┘
💡 Customer Positioning
Always start with Agent Builder. It validates the idea in minutes. If the customer needs more (APIs, workflows, multi-channel), that's when you promote to Studio. This is Microsoft's recommended lifecycle.
🆚 Agent Builder vs Copilot Studio (Full)¶
| Feature | Agent Builder (CS Lite) | Copilot Studio (Full) |
|---|---|---|
| Access | Inside M365 Copilot app | Separate Studio app (copilotstudio.microsoft.com) |
| Coding | None | Low-code + pro-code options |
| Build time | Minutes | Hours to days |
| Knowledge sources | SP, Web, Files, Graph, Email | All above + Dataverse, custom APIs, any data source |
| Skills | Doc gen, code, images, web search | All above + Power Automate, custom actions |
| Actions/tools | Tool Groups (Outlook, SP) | Full connector library (1000+) |
| Workflows | Simple (Copilot Actions) | Complex multi-step with branching |
| Multi-turn dialog | Instructions-based only | Full topic trees with entities and slots |
| Deployment | M365 Copilot + Teams only | M365, Teams, Web, Slack, WhatsApp, custom channels |
| Governance | M365 admin controls | Full Copilot Studio governance + Power Platform admin |
| Cost | Included with M365 Copilot license | May require additional licensing for advanced features |
| Best for | Knowledge assistants, quick prototypes, team-level agents | Enterprise automation, cross-system integration, production agents |
🛣️ Roadmap & What's Coming¶
⚡ Things are moving FAST
Microsoft is shipping Agent Builder updates almost weekly. Features in Preview today may be GA by the time you demo. Always verify in the actual UI.
Recent Additions (Late 2025 – Early 2026)¶
- ✅ Office artifact creation (Word, Excel, PowerPoint) — GA Dec 2025
- ✅ Tool Groups for Outlook and SharePoint
- ✅ Copy to Copilot Studio capability
- ✅ Agent Store for pre-built agents
- ✅ Copilot Actions (scheduled automation)
In Preview (April 2026)¶
- 🔵 Multi-agent orchestration
- 🔵 Human-in-the-loop workflows
- 🔵 Entra Agent Identity
- 🔵 Input widgets (calendar/file pickers)
- 🔵 File groups as knowledge sources
- 🔵 Metadata-filtered SharePoint grounding
- 🔵 Template Hub and Agent Launchpad
Expected (2026 Roadmap)¶
- 🟡 MCP server integration via Copilot Studio
- 🟡 Agent 365 control plane (GA May 2026)
- 🟡 Deeper Graph connector integration
- 🟡 Mobile support
- 🟡 Teams Chat support for custom agents
🧪 Our Demo Agents¶
We're building 10 tiny agents — each showcasing a different capability. Click into each one for the full build guide, architecture, and learnings.
| # | Agent | Primary Feature | Status |
|---|---|---|---|
| A1 | NZ Policy Advisor | 🌐 Web Grounding | ✅ Corp built |
| A2 | Team Wiki Bot | 📂 SharePoint → File Upload | ✅ Corp built |
| A3 | Brand Voice Coach | 📝 Instructions/Persona | ✅ Corp built |
| A4 | Document Generator | 🎨 Office Artifacts | ✅ Corp built |
| A5 | My Comms Assistant | 💬 Teams + Email Knowledge | ✅ Corp built |
| A6 | Meeting Prep Agent | 📅 Multi-Source Combo | ✅ Corp built |
| A7 | M365 Updates Agent | 🔗 Graph Connector (ADO Wiki) | ✅ Corp built |
| A8 | RFP Response Agent | 🌐+📤 Web + File Combo | ✅ Corp built |
| A9 | Email Powerhouse | ⚡ Email-Deep Productivity | ✅ Corp built |
| A10 | 🏆 MAX OUT Agent | 🌟 ALL Features Combined | ✅ Corp built |
| BONUS | Agent Design Assistant | 🛠️ Meta-agent (builds agents) | ✅ Corp built |
| A11 | Fishing Buddy | 🎣 Web Grounding (lifestyle use case) | ✅ Corp built (v2) |
📖 Sources & References¶
| Source | URL |
|---|---|
| Microsoft Learn: Agent Builder | learn.microsoft.com |
| Microsoft Learn: Declarative Agent Architecture | learn.microsoft.com |
| Copilot Studio What's New | learn.microsoft.com |
| M365 Copilot Extensibility | learn.microsoft.com |
| Agent Governance Whitepaper | adoption.microsoft.com |
| Agent365 Samples (GitHub) | github.com/microsoft/Agent365-Samples |
| Copilot Developer Camp | microsoft.github.io/copilot-camp |
| Syskit: Creative Agent Use Cases | syskit.com/blog |
| Ragnar Heil: Agent Builder FAQ | ragnarheil.de |