Agents, Skills, Plugins & Customization MasterclassAgentes, Skills, Plugins & Customização: MasterclassAgentes, Skills, Plugins y Personalización: Masterclass

What Is an AI Agent?

An AI agent is software that perceives, decides, and acts — with autonomy bounded by policy. That definition matters because it draws a clear line between what came before and what enterprises are deploying today.

Yesterday’s chatbot was stateless and reactive: it received a question and returned a single answer. It had no memory across turns, no ability to call tools, and no notion of pursuing a goal. Today’s agent is goal-driven, stateful, and tool-using. It receives a goal, plans a path to reach it, calls external tools, observes the results, and iterates until the goal is met — or escalates to a human when it cannot proceed safely.

Every agent, regardless of platform, is composed of five components:

• Brain — the LLM (GPT-4o, Claude, Llama, Phi) that reasons over input and plans the next step. Picked per task.
• Memory — conversation history, a working scratchpad, a vector store for RAG, and episodic memory across runs. Bounded by policy.
• Tools — functions, APIs, database queries, code execution, workflow triggers. Catalogued and scoped through MCP, plugins, or skills.
• Policy — RBAC, content filters, rate limits, Human-in-the-Loop checkpoints, and audit trails. Non-negotiable.
• Orchestration — the plan-act-observe loop, multi-agent coordination, retries, and timeouts. The platform’s responsibility.

Brain reasons. Memory persists. Tools act on the world. Policy bounds. Orchestration drives the loop.

The Three Platforms

Microsoft offers three distinct platforms for building AI agents. They are complementary, not competitive — most enterprises run all three. The question is which use case lands where.

Copilot Studio targets business users and citizen developers. It provides drag-and-drop topic authoring, native connectors to Microsoft 365, Dataverse, and Power Platform, and a deployment path measured in hours to days. It is the right choice for HR self-service, IT helpdesk automation, and customer-service bots.

GitHub Copilot Agents targets software engineers working inside the IDE and repository. Agent Mode, custom agents defined in .github/agents/, MCP tool integration, and extensions make it the home for code generation, refactoring, PR automation, and test generation.

Azure AI Foundry targets ML engineers and platform teams. The full SDK, a model catalog covering GPT-4o through open-source models, multi-agent orchestration, fine-tuning pipelines, and production-grade observability make it the right foundation for vertical agents with strict SLAs. Time-to-pilot is measured in weeks, but the ceiling is unlimited.

Criteria	Copilot Studio	GitHub Copilot Agents	Azure AI Foundry
Audience	Business / citizen dev	Software engineers	ML / platform teams
Skill level	Low-code, no-code	Code-first, IDE-native	Pro-code, SDK
Best for	HR, IT helpdesk, CS	DevEx, code, PRs	Vertical, multi-agent
Time-to-pilot	Hours to days	Days	Weeks
Governance	Power Platform admin	GitHub policies	Azure RBAC + Content Safety

Agent Types

Four agent types map cleanly onto the platform landscape:

Development agents live in the repository and IDE. GitHub Copilot Agents is the natural home, with custom agents in .github/agents/ and MCP tools for build and ticket integration. Typical tasks: bug-fix, refactor, test generation.

Productivity agents operate inside Microsoft 365. Copilot Studio handles custom topic authoring; Microsoft 365 Copilot Agents deliver embedded experiences in Teams, Outlook, and SharePoint. Typical tasks: meeting notes, email drafts, document search.

Business process agents orchestrate across systems. Copilot Studio handles the conversational layer; Azure AI Foundry takes over when SLAs, multi-agent coordination, or fine-tuning are required. Typical tasks: employee onboarding, approval workflows, claims processing.

Data and analytics agents ground their responses on enterprise data held in Fabric, Synapse, or a custom data warehouse. They require strict prompt engineering, a rigorous evaluation harness, and human sign-off before publishing outputs. Typical tasks: natural-language-to-SQL, insight summaries.

Human-in-the-Loop

More autonomy requires more guardrails, not fewer. An agent that can act on regulated data, sign documents, move money, or change customer records must have explicit human checkpoints. The cost of one wrong autonomous action exceeds the savings of one hundred correct ones.

The HITL decision framework maps every action to one of four quadrants based on impact and reversibility:

• Q1 — Automate: Low impact, high reversibility. Suggest commit messages, draft email replies, summarize meetings, search documents. Run in the agent loop; audit but no checkpoint.
• Q2 — Review: High impact, high reversibility. Open a PR, schedule a meeting, send a draft to a client. Human approves before commit or send.
• Q3 — Confirm: Low impact, low reversibility. Delete a temp file, archive a ticket, close a conversation. Single confirmation step.
• Q4 — Dual Control: High impact, low reversibility. Sign a contract, transfer funds, change RBAC, delete production data, file taxes. Two humans approve; audit plus signed evidence required.

Three implementation patterns cover most cases. Approve-before has the agent propose and wait at a checkpoint before acting — correct for PRs, drafts, and recommendations. Sample-and-review lets the agent act at high volume while a random sample is audited and KPIs trigger escalation — correct for tagging, classification, and routing. Dual-control requires an initiator and a separate approver with cryptographic signatures on both — correct for finance, legal, and production data.

Case Study: Vehicle Leasing

Two real operational problems illustrate multi-agent design in the vehicle leasing and sales industry.

Traffic-fine driver indication has a 15-day legal deadline. Miss it, and the leasing company pays the fine plus penalty. The manual process today is: mailroom receives a paper notice, an operator scans and transcribes it into Excel, emails the customer who leased the car, and hopes for a timely reply. Deadlines slip regularly.

The agent-led flow runs four specialized agents on Azure AI Foundry. An ingestion agent (Document Intelligence) extracts plate number, infraction date, code, value, and deadline from the scanned notice. A lookup agent cross-references SAP and Salesforce to identify the active lessee. A notification agent sends a portal link, SMS, and email — with HITL review on the notification template by the operations team. A tracking agent escalates automatically on day 7, sends a final ping on day 12, and routes to a paralegal on day 14.

Tax compliance for NF-e faces a different failure mode: one wrong CFOP code blocks invoices for an entire month. Every contract closure, lease termination, and inter-branch vehicle transfer must produce a valid Nota Fiscal Eletrônica conforming to RICMS rules that vary by state and product category (NCM). Manual issuance breaks at scale.

The compliance flow runs three agents triggered by an ERP event (SAP, Oracle, or TOTVS). A RAG-based tax classifier uses Azure AI Search over RICMS legislation and the NCM table to select the correct CFOP, CST, and ICMS rate. An NF-e issuer agent calls the SEFAZ NF-e 4.0 API, handles retries, contingency mode, and cancellation flow. A bookkeeping agent runs a daily reconciliation — SPED, EFD-Contribuições, in/out ledgers, and the ICMS-ST report. Monthly tax filing carries dual-control HITL with CFO sign-off.

Vertical Examples

The same multi-agent pattern — ingest, classify, act, escalate — transfers across industries:

• Finance: A document agent reads onboarding files, a risk-scoring agent matches sanctions lists, and an escalation agent routes to a compliance officer with dual-control on KYC/AML approval.
• Healthcare: A clinical-extraction agent plus a payer-rules RAG layer, with HITL involving the physician before prior-authorization submission. Authorization time drops from days to hours; all decisions are logged for audit.
• Retail: A triage agent classifies a return reason, a refund agent runs the policy check, and a logistics agent books pickup. HITL applies only to exceptions above the SKU value threshold.
• Manufacturing: A telemetry agent monitors IoT signals, a diagnostics agent proposes work orders, and a scheduling agent slots technicians. The plant manager approves any work order longer than four hours.

Architecture Principles

Production-grade agent programs are built in six layers, from the ground up: Platform & Governance (identity, RBAC, Content Safety, audit, observability, FinOps — the same controls as the rest of Azure), Models (Foundry catalog, picked per task, fine-tuned where justified), Context & RAG (vector store, semantic retrieval, identity-scoped access), Tools & MCP (catalogued, versioned, and scoped MCP servers and plugins), Orchestration (multi-agent runtime, planning, retries, HITL checkpoints, evaluation), and Surface (Teams, Outlook, IDE, or custom app — where users meet the agent).

Five principles govern every Microsoft AI agent program:

1. Start narrow. One use case, one agent, one team. Generalize only after value is proven.
2. Reuse Azure governance. Identity, RBAC, Content Safety, Purview — the same controls as the rest of the stack, not parallel ones.
3. Catalogue every tool. MCP server registry, scoped tokens, audit per call. No agent talks to a random API.
4. Evaluate before deploy. Golden test set per agent, regression on every model swap, latency budget per loop.
5. Plan the deprecation. Version every prompt, log every change, keep the rollback — agents drift like models do.

Deployment Roadmap

The path from first pilot to a production-grade agent fleet runs five phases over twelve months:

Phase 1 — Discover (Month 1): Map use cases against value and effort, get stakeholder sign-off, select the pilot candidate.

Phase 2 — Pilot (Months 2–3): Deploy one agent with one team. Run a closed-loop evaluation before any broader access.

Phase 3 — Harden (Months 4–6): Implement RBAC and audit trails, set SLA targets and cost caps, wire HITL checkpoints into every relevant action.

Phase 4 — Scale (Months 7–9): Build the tool registry, introduce multi-agent orchestration, open self-service onboarding for new teams.

Phase 5 — Operate (Months 10–12+): Establish FinOps and drift monitoring, run model-swap drills, conduct quarterly reviews against the original value targets.

Each phase has explicit exit criteria. Piloting without hardening produces a demo. Scaling without operating produces debt.