Meet the Agent Academy Hackathon Winners

From May 12 to June 2, 2026, builders from around the world put their Agent Academy learning to the test by designing, building, and shipping real, working AI agents with Copilot Studio. After two weeks of judging, we’re thrilled to announce the Agent Academy Hackathon winners.

🏆 The Winners

Recruit Track

The Recruit track is geared toward exploring the fundamentals of agent building and is optimized for first-time agent builders.

🥇 First place: Performance Development Assistant

Built by @Ateina

What they built

After a performance review, employees often know what to improve but have no clear path forward and they’re unaware of the learning materials and mentors that already exist inside their own company. This Copilot Studio agent turns review feedback into a structured development roadmap. On startup it pulls the company’s skills list from SharePoint, then loops through each skill the employee wants to grow: it surfaces a Microsoft Learn answer, looks up internal mentors by expertise from a SharePoint list, and returns relevant internal materials as an adaptive card. If a restricted learning resource is requested, a Power Automate flow routes a manager-approval email. It earned the track’s top technical-execution score (9/10).

Agent Academy concepts on display:

1. Copilot Studio agent grounded in Microsoft Learn as a knowledge source

2. Three SharePoint lists (Skills, Mentors, Materials) as the internal knowledge backbone

3. Adaptive Cards built with Power Fx + SharePoint data

4. Power Automate approval flow triggered from card input

5. Custom topics, generative answers, and a loop pattern over the skills list

🔗 View the repo ▶ Watch the demo

🥈 Second place: Meeting Tasks Agent

Built by @hutten358

What they built

Meeting Tasks Agent tackles the gap between a meeting ending and the follow-up work actually getting done. The agent takes a meeting transcript, supplied as a Word document, and turns the discussion into a clear set of tasks the team can act on. Judges liked the practicality of the scenario and the clean transcript-to-tasks flow.

Agent Academy concepts on display:

1. Copilot Studio agent that converts a meeting transcript into actionable tasks

2. Document-grounded input (Word transcript) → structured task output

3. Recruit-track fundamentals, clear use case and a working, demoable agent experience

🔗 View the repo ▶ Watch the demo

🥉 Third place: Conversational AI Agent for Vehicle Insurance Self-Service Portal

Built by @anupamac985

What they built

This custom Copilot Studio agent takes the two highest-volume, rules-based contact-centre requests, submitting a vehicle-insurance claim and cancelling a policy, and turns them into guided, validated self-service journeys available around the clock. It recognises intent, runs a multi-turn dialogue, makes decisions from live data (policy status, eligibility, refund calculation), and writes results back to Dataverse. A reusable identity-verification step is shared by both journeys, with responsible-AI guardrails: escalation to a human after two failed identity checks and an explicit confirmation before any irreversible action.

Agent Academy concepts on display:

1. Custom Copilot Studio agent grounded in structured Dataverse data

2. Four intent-triggered topics, including a reusable identity-verification topic

3. Power Automate flows for claim creation, eligibility checks, and cancellation

4. Responsible-AI guardrails: human escalation and confirmation gates

5. Global variables passed across topics for a continuous experience

🔗 View the repo ▶ Watch the demo

Operative Track

For builders creating more advanced agents with concepts like orchestrating multi-agent solutions, prompts, Agent Flows, security, etc.

🥇 First place: VendorGuard — Autonomous Vendor Contract Compliance System

Built @experienceswithanishh

What they built

The single highest-scoring entry of the entire hackathon. VendorGuard is a fully autonomous, multi-agent vendor-contract compliance system for procurement teams. When a contract PDF lands in an inbox, a Power Automate flow triggers, creates a Dataverse record, extracts the text with an AI prompt, posts a Teams notification, and hands off to a Copilot Studio orchestrator. The orchestrator delegates to four specialist agents that score the contract against 15 compliance rules across five dimensions (Commercial, Legal, Data Privacy, SLA & Performance, Regulatory), rating each clause Red/Amber/Green. It then stores a compliance review report and makes every contract queryable in natural language. It takes you from email to report in under two minutes, with the rulebook stored in Dataverse so it can change without touching the agent.

Agent Academy concepts on display:

1. Multi-agent hub-and-spoke orchestration (orchestrator + 4 specialist agents)

2. Event-triggered autonomous pipeline (email arrival → Power Automate)

3. AI Builder prompt for PDF content extraction

4. Microsoft Dataverse MCP Server for intelligent data operations

5. Dataverse-grounded 15-rule compliance rulebook + stored review report

6. Teams notifications and AI-safety / scope-redirection handling

🔗 View the repo ▶ Watch the demo

🥈 Second place: Engagement Hub

Built by @leila-marspooner

What they built

Engagement Hub helps engagement managers and delivery leads at consultancies handle the fragmented stream of client documents, SOWs, and service requests. A parent agent analyses a submitted client document via a Power Automate flow that extracts and assesses the text with an AI prompt, then saves a structured analysis with risk flags, a confidence score, human-review status, and a correlation ID. The user can create a duplicate-safe Engagement Request, after which a Delivery Prep child agent generates a discovery brief and a Word handoff document saved to SharePoint and linked back to Dataverse. If human review is needed, a Dataverse-triggered flow posts a Teams adaptive card to the review channel.

Agent Academy concepts on display:

1. Parent + connected + child-agent orchestration with controlled topics

2. Multi-modal AI prompt with structured JSON output, grounded in Dataverse

3. Automated human-in-the-loop: Dataverse-triggered Teams adaptive-card review

4. Discovery-brief + Word handoff doc generation to SharePoint

5. Duplicate-safe record creation and audit logging (correlation IDs)

🔗 View the repo ▶ Watch the demo

🥉 Third place: FrostByte AI Advisor

Built by @JBearCode

What they built

FrostByte AI Advisor gives ‘Brain Freeze’ ice-cream store managers the data and actions they need to run the shop. It’s designed to solve 11 manager problems: analysing sales by item, flavour and topping; forecasting demand; drafting restock orders; running pricing scenarios; analysing and extending promotions; and even suggesting promotions by reading competitor flyer PDFs. A parent agent handles read/analysis while a connected child agent performs all write operations to Dataverse via the Dataverse MCP server. A multimodal prompt ingests competitor promo PDFs, and AI reasoning grounded in live Dataverse data outputs reliable JSON that is patched back using Code Interpreter. It ships as an importable solution and was tuned to run on Claude Sonnet for the best quality/speed/cost balance.

Agent Academy concepts on display:

1. Parent + connected child agent (reads vs. Dataverse writes)

2. Dataverse MCP server for create operations

3. Multimodal prompt extracting competitor promo flyer PDFs

4. Dataverse-grounded prompts patched back via Code Interpreter

5. Event triggers (scheduled + SharePoint) and AI-safety self-identification

6. User-feedback collection via adaptive card

🔗 View the repo ▶ Watch the demo

Special Ops Track

For advanced patterns and specialized scenarios.

🥇 First place: SprintForge

Built by @Shrusti13

What they built

SprintForge is an AI scrum assistant that converts call transcripts and rough planning notes into structured sprint plans and creates the work items automatically. A Scrum Master simply forwards an email (with optional Teams transcripts, PDFs, or Word attachments) to SprintForge. A Copilot Studio agent reads the body, extracts sprint goals, epics, stories, tasks, owners, and blockers, and identifies the Azure/Microsoft services mentioned. It then calls the Microsoft Learn MCP server to fetch authoritative references and enrich each story with acceptance criteria and implementation notes, before using a Jira MCP server to create the issues directly in Jira Cloud. The team reports up to a ~70% reduction in sprint-prep time, a unanimous favourite among the primary judges.

Agent Academy concepts on display:

1. Copilot Studio agent combining extraction, enrichment, and automation

2. MCP tool chaining across two servers (Microsoft Docs MCP + Jira MCP)

3. Microsoft Learn grounding to enrich every story with best practices

4. Multi-source ingestion (email body + attachments + transcripts)

5. Structured output (Sprint Plan JSON → Jira issues)

🔗 View the repo ▶ Watch the demo

🥈 Second place: Calamity Agent — Weather & Fire Alerts

Built by @tagr

What they built

Calamity Agent is a disaster-preparedness assistant. Ask it about hazards near a location (“What alerts are near Seattle?”) and it returns current National Weather Service alerts, near-real-time fire detections from NASA’s FIRMS (VIIRS) data, and maps of the affected areas, then recommends preparedness guidance grounded in FEMA/Ready.gov documents. Under the hood, a Copilot Studio agent (running Claude Sonnet 4.6) calls a custom MCP server over Streamable HTTP; the server calls a .NET Aspire ‘Calamity’ API that geocodes via Azure Maps, pulls NWS zones and active alerts, fetches VIIRS fire detections, and renders static maps. Judges highlighted the AI-safety-first design and the sheer amount of real engineering behind it.

Agent Academy concepts on display:

1. Custom MCP server (Node.js/TypeScript, MCP Streamable HTTP) exposing five tools

2. MCP tool chaining: geocode → alerts → map → fire

3. Azure backend (.NET Aspire API on Azure Container Apps, deployed via azd)

4. External-system integration: NWS, NASA FIRMS, Azure Maps

5. Knowledge grounding in FEMA / Ready.gov docs; adaptive cards for location prompts

6. Zod input-schema validation for structured tool inputs

🔗 View the repo ▶ Watch the demo

🥉 Third place: Warehouse Picking Agent

Built by @granjan7779

What they built

This entry is a custom MCP server that connects an agent to Microsoft Dynamics 365 for warehouse picking operations. The server exposes warehouse-work tools over HTTP/MCP: read open pick lines (outbound sales-order picking work) and confirm warehouse work orders back into D365, passing real warehouse fields such as work-line IDs, target license plate, quantity, and work type. The tool schema is driven from Azure AI Foundry, and the same server plugs into both Copilot Studio and AI Foundry. Judges saw a strong business case for replacing specialised warehouse-picking devices with conversational AI.

Agent Academy concepts on display:

1. Custom MCP server (Node.js) over HTTP/REST

2. Dynamics 365 Warehouse Management integration (pick lines, work confirmation, license plates)

3. Plugs into both Copilot Studio and Azure AI Foundry

4. Structured tool inputs with required-parameter validation

5. D365 / Azure AD authentication module; deployable to Azure

🔗 View the repo ▶ Watch the demo

Cowork Collective Track

For builders delegating real work to Copilot Cowork.

🥇 First place: Copilot Cowork Autonomous ITSM Platform on Microsoft 365

Built by @ninihen1

What they built

An ambitious, end-to-end ‘Autonomous IT Service Desk for Small & Medium Business’ that runs entirely on Microsoft 365. A user submits a ticket through a SharePoint-hosted portal; a Power Automate flow hands it to a Copilot Studio triage agent that classifies it, checks the knowledge base, and returns a read-only outcome (resolve from KB, ask for detail, or propose an action). Any tenant-changing action pauses for a manager’s one-tap approval via a Teams adaptive card, after which a narrowly-scoped service principal executes the change through Microsoft Graph and logs an audit record. Notably, the entire stack, SharePoint lists, Power Automate flows, and the SPFx React portal, was built and maintained by AI agents (Copilot Cowork with a Flow Studio MCP, plus an IDE agent) under one person’s supervision.

Agent Academy concepts on display:

1. Copilot Cowork delegated to build & maintain Power Automate flows (Flow Studio MCP)

2. Copilot Studio triage agent with read-only outcomes

3. Human-in-the-loop: one-tap Teams adaptive-card manager approval

4. Least-privilege execution via a single scoped service principal + Microsoft Graph

5. Deep M365 integration: SharePoint (SPFx React portal), Power Automate, Teams

🔗 View the repo ▶ Watch the demo

🥈 Second place: Client Kick-off Skill

Built by @appieschot

What they built

A focused Microsoft 365 Copilot Cowork skill that turns a signed proposal or engagement deck into a ready-to-run client kickoff — meeting, agenda, Word document, and a Finance intake email — in one approval-gated conversation. It runs in two gated phases: Phase 1 (read-only) finds and reads the attached deck (PPTX/PDF/DOCX), extracts a project summary, deliverables and tasks, and an attendee list resolved against the directory, then stops. Phase 2 only starts after the user confirms attendees, and produces four outputs — each executed only on explicit per-item approval: a Teams kickoff meeting in the soonest common slot, a time-boxed agenda, the agenda saved as a Word doc in OneDrive, and a structured Finance intake email. Judges singled out how clearly the presentation conveyed the scenario and its value.

Agent Academy concepts on display:

1. Single Copilot Cowork custom skill (auto-discovered from OneDrive)

2. Two-phase, approval-gated workflow with a hard stop between phases

3. Per-item human-in-the-loop approval (no silent sends)

4. Directory/people resolution + Outlook Calendar + Word + OneDrive + mail orchestration

5. Grounded, anti-fabrication generation (unknowns flagged as placeholders)

🔗 View the repo ▶ Watch the demo

🥉 Third place: Team Yaito — A Multi-Agent PMO on Copilot Cowork

Built by @Jirawat-Yaito

What they built

Team Yaito is an ambitious multi-agent Project Management Office (PMO) built entirely on Copilot Cowork. A single orchestrator agent (Yaito) coordinates a team of specialist agents across five PMO functions — plan, architect, govern, measure, and scale — to help run enterprise program delivery, with a human accountable for every consequential action. It targets the manual, fragmented way enterprise PMOs run today — status decks, scattered chat threads, and one-off spreadsheets that consume senior PM time while leadership lacks a real-time view of delivery health. The orchestrator decomposes a natural-language request, delegates to the right specialist agent (which acts on live tenant data through MCP/Graph tools across Outlook, Teams, SharePoint, and Calendar), and routes every write action through a human approval step before it commits. The submission frames it around a real 30-day HR-chatbot migration heading toward a mid-June go-live.

Agent Academy concepts on display:

1. Copilot Cowork as a multi-agent runtime (orchestrator + specialist agents)

2. Specialist agents organised across five PMO functions (plan, architect, govern, measure, scale)

3. Delegated work via natural-language task decomposition

4. Microsoft Graph integration (Outlook, Teams, SharePoint, Calendar) through MCP tools

5. Human-in-the-loop approval before any write action

6. Agents declared as Markdown + YAML definition files

🔗 View the repo ▶ Watch the demo

What every submission needed

1. A working agent built using at least one Microsoft product — Copilot Studio, M365 Copilot, or Copilot Cowork.

2. A demo video no longer than five minutes showcasing the solution.

3. An architecture overview showing components, data flow, and integration points.

How Entries Were Judged

Every qualifying entry was scored 1–10 on six criteria and combined into a weighted total.

• Accuracy & Relevance – 25%

• Technical Execution – 25%

• Creativity & Originality – 15%

• User Experience & Presentation – 15%

• Reliability & Safety – 10%

• Use Case Impact – 10%

Thank You!

Congratulations to all of our winners, and thank you to every builder who participated. The creativity, technical depth, and real-world impact across all four tracks was genuinely inspiring and your feedback will directly shape where the Agent Academy curriculum goes next.

Want to explore more and start building your own agent? Visit https://aka.ms/agent-academy to get started!