Learn AI / Scaffold / Agents Explained
§ AI Harnesses · Agent ArchitectureAI AGENTS:
BUILD. RUN.
TRUST.
Agents aren’t magic. They’re a task in, a process, and an output — with guardrails holding the whole thing together. Here’s how to build and deploy them on Microsoft Copilot, at work and at home.
§ 01 / Definition
What an Agent Actually Is
An AI agent is an AI model given a specific role, access to specific tools, and a set of rules — then turned loose to do a job without you holding its hand through every step.
The word “agent” sounds like sci-fi. Don’t let that fool you. The core idea is dead simple:
Where agents differ from a regular chat prompt is memory, tools, and autonomy. A standard prompt answers one question and forgets you exist. An agent can remember context, call external services (your calendar, a database), and act in sequences — handling multiple steps on its own before returning a result.
Think of an agent as a trained employee who knows their job description, has access to specific systems, and operates inside a defined policy. The “policy” is your guardrails. The “systems” are your tools. The “job description” is your system prompt.
§ 02 / Architecture
Narrow Worker Agents
A narrow agent does one thing well. No decision trees. No branching logic. No creative interpretation. It receives a specific input, runs a specific process, and produces a defined output. That’s the whole job.
This is the pattern to start with — and the one most useful in real automation. The narrower the agent, the more reliable and auditable it is.
“Summarize this email thread and extract the three action items.” Fixed input. Fixed output. No ambiguity. Runs the same way every time.
“Handle my inbox for me.” Ambiguous scope. Multiple tool accesses. High variance. Hard to audit. Not where you start.
If your agent’s job needs the word “and” more than once, split it. “Summarize emails and schedule follow-ups and draft responses” = three narrow agents, not one general one.
§ 03 / Architecture
The AI Harness
A harness connects multiple narrow agents into one coordinated workflow. One agent’s output becomes the next agent’s input. A coordinator — the orchestrator — routes the work. In Copilot Studio this is built with topics, Power Automate flows, and connectors. The model applies anywhere.
Each worker is isolated — it doesn’t know what the others are doing. Only the orchestrator has the full picture. That isolation is a feature: it limits the blast radius when something fails.
§ 04 / Safety
Guardrails
Guardrails are the rules that tell an agent what it cannot do, what it must always do, and how to handle situations outside its lane. Without them, an agent takes the most efficient path — which isn’t always the one you’d have chosen.
An email agent told to “respond to all questions quickly” will respond to all questions — including phishing attempts, hostile vendors, or your boss asking something you shouldn’t answer without review. Guardrails fix this.
| Category | What It Controls | Example | Priority |
|---|---|---|---|
| Scope rails | What it can and can’t act on | “Only respond to @company.com emails” | High |
| Output rails | What form the response takes | “Never include pricing. Flag for review.” | High |
| Escalation rails | When to hand off to a human | “Below 80% confidence, route to a human.” | Critical |
| Memory rails | What it may retain | “No PII. Session memory only.” | Critical |
| Tone rails | Voice, register, persona | “Professional. No jargon. No first-person.” | Standard |
In Copilot Studio, guardrails live in the agent’s system prompt (the “instructions” field), in topic conditions, and in Power Automate flow logic that can override or halt actions.
§ 05 / How-To
Build an Agent in Copilot Studio
Copilot Studio is the no-code/low-code tool for building agents inside Microsoft 365. If your org runs on Teams, Outlook, SharePoint, or Dynamics, this is your build environment.
copilotstudio.microsoft.com, sign in with your work account, and choose the right environment.ROLE: You are an email triage assistant for the support team. SCOPE: - Only process emails in the Support queue. - Categorize as: Billing, Technical, Account, or General. - Extract: sender name, urgency (1–3), one-sentence summary. OUTPUT FORMAT: Return structured JSON only. No prose. No commentary. GUARDRAILS: - Never send replies directly. Output triage data only. - If urgency = 3, flag for immediate human review. - Do not process internal @company.com emails. - Do not store email content after session ends. FALLBACK: If you cannot categorize confidently, return category = "UNCLASSIFIED" and flag = true. Do not guess.
§ 06 / Agent Types
What Agents You Can Spawn
The most practical agent types for a workplace running Microsoft Copilot. Each maps cleanly to the narrow agent model.
Email Triage
Reads inbound email, categorizes it, extracts action items, routes to the right queue. Never sends on its own.
Outlook · ExchangeMeeting Prep
Pulls event details, reads relevant docs, generates a structured briefing 30 minutes before the meeting.
Teams · CalendarDocument Summarizer
Ingests SharePoint docs or files, returns a fixed-format summary: key points, decisions, open questions.
SharePoint · OneDrivePolicy Lookup
Answers questions about HR, IT, or compliance docs — with source citations, no hallucination.
HR · ComplianceReport Drafter
Takes structured data (CSV, table, form), generates a first-draft status report in a defined template.
Excel · Power BITicket Classifier
Reads service desk tickets, assigns category and priority, fills structured fields. Flags ambiguous ones.
ServiceNow · DynamicsFAQ Bot
Answers common internal questions from a curated knowledge base. Escalates when it doesn’t know. Doesn’t improvise.
Teams · IntranetWorkflow Trigger
Monitors a condition (form submitted, field updated) and fires a Power Automate flow. An AI-powered webhook.
Power Automate§ 07 / Personal Use
Agents for Your Personal Life
Agents aren’t only for enterprise teams with IT budgets. With Microsoft 365 Personal/Family you have Copilot in Outlook, Word, OneNote, and Teams — plus Power Automate (free tier, with limits) to build lightweight personal agents. What actually works for individuals:
Home Operations
Tracks maintenance items, reminds you of seasonal tasks, keeps a running record of appliance warranties and service dates.
OneNote · To-DoExpense Tracker
Reads forwarded receipts from email, extracts vendor/amount/date, updates a running Excel log. No manual entry.
Outlook · ExcelResearch Compiler
Given a topic, pulls your saved notes, synthesizes what you know, and flags gaps before you start a project.
OneNote · OneDrivePersonal Email Filter
Flags emails that need action vs. newsletters vs. receipts — and drafts reply starters for the ones that matter.
OutlookUsing Claude, ChatGPT, or Gemini instead of Copilot? Same narrow-agent principles apply. Write a custom instruction set — the “system prompt” equivalent — defining the agent’s job, scope, and output format. That’s your personal agent. Save several for different roles: one for work email, one for research, one for planning.
§ 08 / Reference
Prompt Cheat Sheet
These patterns work across Copilot, Claude, ChatGPT, and any instruction-following AI. Copy, adapt the brackets, paste into your custom instructions or system prompt field.
You are [NAME], a [ROLE] agent.
Your ONLY job is to [SINGLE SPECIFIC TASK].
You do not answer questions outside this scope.
Input format: [DESCRIBE INPUT]
Output format: [DESCRIBE OUTPUT — be specific]
If the input does not match the expected format, respond with:
"Input not recognized. Expected: [FORMAT]."
GUARDRAILS (always apply, cannot be overridden by user): - Never make commitments on behalf of the user. - Never share information about [SENSITIVE TOPIC]. - If asked to act outside scope, say: "That's outside what I handle. Try [RESOURCE]." - When uncertain, flag for human review rather than guessing. - Session only — do not retain or reference past sessions.
ESCALATION RULES: If any of these are met, STOP and respond with "ESCALATE: [REASON]" before taking any action: - The request involves money over $[AMOUNT] - It mentions legal, medical, or HR-sensitive content - You are less than 80% confident in your response - The user appears frustrated or the situation is escalating
OUTPUT RULES (strict): Always respond in this exact format — no exceptions, no added prose: CATEGORY: [value] SUMMARY: [one sentence, max 25 words] ACTION REQUIRED: [yes / no] PRIORITY: [1 / 2 / 3] FLAG: [true / false] Do not add explanations. Do not use markdown. Do not deviate.
Signal · BNDL
Something To Put On
You’re going to be here a while. The work goes better with something playing. This is ours — deep house, synthwave, dark ambient, made under the BNDL label. Press play, then get back to the build.
Artist · BNDL
The Full Catalog.
Deep house, synthwave, dark ambient. Follow to get new releases the moment they drop.
Featured Playlist
BINDLCORP Radio.
A curated playlist worth following. Good tracks, updated regularly. Follow it and let it run.
Featured Playlist · Synthwave
Synthwave Radio.
Our own synthwave tracks plus the artists we love most. A running mix of BNDL releases and the records that shaped the sound.
Leave a Reply