More Amplifier Stories
Experimental
M365 + Amplifier Hackathon

We Severely Underestimate
What AI Agents Can Do

A story about discovery, surprise, and the future of human-AI collaboration

January 2026
The Challenge

Can Amplifier build its
own M365 integration?

No existing code. No documentation. No hand-holding.
Just a test tenant and a simple instruction.

The Setup

One prompt. One hour. Zero prep.

"I just got a M365 test tenant for you! You don't have the code/tools/etc to do it, so you'll have to start by building yourself those things."
The only instruction given to Amplifier
What Users Provided
Tenant ID, Client ID, and copy-pasted permissions. That's it.
What Users Did NOT Provide
Architecture decisions, code, API knowledge, authentication logic, error handling...
What Happened

Amplifier took over

Designed the architecture
Decided on a Python module structure with 11+ source files
Built 3-tier authentication
ClientSecretCredential → DeviceCodeCredential → DefaultAzureCredential fallback chain
Created operation modules
email_send, email_read, calendar_list_events, user_list, sharepoint_search, teams_send_message
Wrote specialized agents
m365-admin.md and m365-productivity.md with distinct personas
The Turning Point

Autonomous adaptation

When the user provided different permissions than expected, Amplifier didn't ask what to do. It pivoted.

"Actually, I need to add Teams operations! The current implementation has Email/Calendar but the permissions granted are for Teams/Chat/Files/Tasks."
Amplifier, noticing the mismatch and adapting without being asked
"Actually, I realized you can do this for me... let me know when you have a device code for me to try"
User's realization that they could step back and let Amplifier lead
Key Insight

Zero technical depth
required

Participants just followed instructions Amplifier gave them.
Navigate to Azure portal. Grant these permissions. Click here. Paste this.

0
Lines of code written by users
0
API docs consulted
Repeatable
100%

Success rate

Based on participant accounts

Week 1
Small group mini-hackathon
Everyone succeeded
Week 2
Much larger group
Same result across all participants
Extended Exploration

Week two: Going deeper

With working M365 integration, participants explored what's actually possible.

5 Agent Users
Created in M365: amplifier-agent-1 through -5, each with Teams license
"AI Agents" Team
8 specialized channels: Handoffs, Task Coordination, Alerts, Decisions, File Exchange, Logs
Polling Daemon
Python service that watches for @mentions and routes to Amplifier for responses
A New Paradigm

Agent users are
real M365 citizens

Not API consumers. Actual team members.

The Breakthrough

The "IT WORKED" moment

Processing Teams message mentioning Amplifier Agent 1 from MOD Administrator Created new session teams:acad129f-a50b-4a3a... HTTP Request: POST https://api.anthropic.com/v1/messages "HTTP/1.1 200 OK" Sent Teams reply successfully Agent Amplifier Agent 1 responded to Teams message
📧 Checking for new emails for Amplifier Agent 1... Found 1 unread email(s) Processing email from MOD Administrator Agent generated response (length: 847) ✅ Reply sent successfully!
Self-Correction

Autonomous problem solving

When Teams split "@Amplifier Agent 1" into 3 separate mentions, Amplifier diagnosed and fixed it without being asked.

"I see the issue! Teams split your @mention into 3 separate mentions... The daemon needs to check the user ID in the mention, not just the display name."
Amplifier, diagnosing a subtle Teams API behavior
Then it just... fixed it
Modified the code to use user ID matching. Tested it. Confirmed it worked. Moved on.
What's Next

The future is multi-agent

Personal Uplevel
My agent does work for me autonomously. Monitors my inbox. Drafts responses. Schedules meetings. Summarizes threads I missed.
Agent Proxies
"Have my agent talk to Sam's agent" - they might collaborate via Teams messages, shared docs, or structured handoffs.
Device Coordination
Amplifier instances on different machines (home, office, server) coordinating work via M365 as the communication backbone.
Development Velocity

What Amplifier built autonomously

11+
Source Files
6
API Operations
3
Auth Methods
5
Agent Users
8
Teams Channels
1
Polling Daemon
Sources

Research Methodology

Data as of: February 20, 2026

Feature status: Experimental

Research performed:

Gaps: No formal "hackathon" event found in git history. Narrative based on internal exploration sessions, January 2026. Success rate (100%) is from participant accounts, not formal measurement. Channel count (8) not independently verified.

Primary contributors: Sam Schillace (platform + collab), Marc Goodner (platform PRs)

The Takeaway

We severely underestimate
what AI agents can do

The bottleneck isn't the AI's capability.
It's our imagination about what to ask for.

M365 + Amplifier Hackathon • January 2026
← More Amplifier Stories More Amplifier Stories