Agents That Clock In

Amplifier as co-developer through the M365 auth gauntlet

The Goal

Several Amplifier instances, collaborating on real work inside Microsoft 365

A base agent plus four named personas, coordinating over a shared SharePoint message board. Getting autonomous agents past M365 auth is the hard part everyone underestimates.

First, proof the collaboration is real — then the wall it had to get through.

The ANDMulti-agent collaboration on real M365 surfaces — Teams, SharePoint, and more.
The BUTWiring agents into M365 is an auth / permissions / MFA / licensing gauntlet.

Proof — It Shipped Fast

The five-persona Banzai bundle’s core was built in a single day

amplifier-bundle-m365-collab (Sam Schillace) defines five persona files — a base agent plus four Hong Kong Cavaliers — collaborating over a SharePoint message board. Its core multi-agent system was built in one session on 2026-01-20.

Working collaboration exists. So why is M365 the hard part?

5

persona files: base agent + Buckaroo Banzai, New Jersey, Perfect Tommy, Reno Nevada

~5.5h

core built in one session, 14:59 → 20:30 on 2026-01-20

3,200

lines added / 74 deleted across 9 commits

1

author — Sam Schillace (100% of commits)

The Gauntlet

An autonomous agent can’t just log in — M365 throws a wall of Azure AD failures

The amplifier-m365-platform repo catalogs the real errors an agent hits, including MFA-required and consent. These aren’t hypotheticals — they’re counted straight out of the repo’s troubleshooting docs.

Clearing this wall needs a real backbone. Meet the platform.

AADSTS50076 ×8Multi-factor authentication required to access the resource.
AADSTS65001 ×3Consent not granted for the application.
AADSTS50034 · AADSTS7000218 · AADSTS90002 · AADSTS500011More cataloged Azure AD failures in the same repo.

The Backbone

amplifier-m365-platform is the auth and tooling backbone

Co-authored ~76% Sam Schillace / ~24% Marc Goodner. It ships an auth decision tree, scripts to create agent users and assign licenses, and modules wrapping Microsoft Graph surfaces — Teams, SharePoint, Dataverse, and more.

With the backbone in place, one auth decision does the heavy lifting.

16

tool / hook / orchestrator module directories under src/

37

commits: 28 Schillace (~76%) + 9 Goodner (~24%)

2

scripts: create-agent-users.py & assign-licenses.py

4

auth flows in the decision tree: ROPC, OBO, Device Code, Client Credentials

The Mechanism

The decision tree picks ROPC so an agent can act AS a specific user

For “autonomous agent as user,” the auth decision tree selects ROPC — Resource Owner Password Credentials — yielding a delegated token scoped to a specified user. One deliberate choice turns “agents can’t log in” into “agents act as users.”

And Amplifier didn’t just recommend that — it helped build it.

# auth-decision-tree.md
| Scenario                 | Flow  | Token     |
| Autonomous agent as user | ROPC  | Delegated |
| On-behalf-of             | OBO   | Delegated |
| Device Code              | ...   | Delegated |
| Client Credentials       | ...   | App-only  |

The Turn

Amplifier didn’t just advise — it acted as co-developer

The validation work landed as real commits in the platform repo — footered to credit Amplifier itself as a contributor, not just a tool that gave advice.

Generated with Amplifier
Co-Authored-By: Amplifier

Describe the vision, let Amplifier build — then prove it against a real tenant.

The Payoff

Validated end-to-end against a real M365 tenant — the gauntlet, cleared

On 2026-01-23, commit a864f28 recorded the platform’s end-to-end validation with real M365 credentials against an actual tenant: 4 of 5 checks PASS, 1 PARTIAL — carrying the “Generated with Amplifier” footer.

So what’s the pattern worth keeping?

4/5

end-to-end checks PASS against a real tenant

1

PARTIAL result in the same E2E run

01-23

2026 — validation commit a864f28

The Pattern

Describe the vision, let Amplifier architect-build-test as co-developer

A multi-agent M365 integration that normally drags on came together in about a day: the Banzai bundle’s core built on 2026-01-20, the platform E2E-validated by 2026-01-23. Bring the intent; let Amplifier clear the gauntlet.

Multi-agent collaboration on real M365 surfaces — from vision to validated in days.

Sources

Research Methodology

Data as of: commits through 2026-01-23. Feature status: amplifier-m365-platform E2E-validated 4/5 PASS against a real tenant (SUPPORTED).

Repositories: ramparte/amplifier-bundle-m365-collab (Sam Schillace) and ramparte/amplifier-m365-platform (Sam Schillace ~76% / Marc Goodner ~24%). Diego Colombo (GitHub colombod, id 375556) is the featured deck persona, not a committer.

Commands run:

ls amplifier-bundle-m365-collab/personas/ — 5 persona files (base + 4 Cavaliers)
git -C amplifier-bundle-m365-collab log --format='%ci %s' — core built 2026-01-20 (14:59 → 20:30); numstat 3,200 added / 74 deleted
grep -rhon 'AADSTS[0-9]*' amplifier-m365-platform | sort | uniq -c — AADSTS50076 ×8, AADSTS65001 ×3, and others
sed -n '1,20p' .../auth-decision-tree.md — ROPC = “Autonomous agent as user,” delegated token
ls .../m365-tenant-setup/scripts/ — create-agent-users.py, assign-licenses.py; 16 tool/hook module dirs under src/
git -C amplifier-m365-platform show --stat a864f28 — E2E 4/5 PASS, 1 PARTIAL; “Generated with Amplifier” footer
gh api users/colombod --jq '{login,id,name}' — colombod, id 375556, Diego Colombo

Gaps / not claimed: No Planner handoff run or named agent-users were found in-repo; Planner is a documented capability only. The “~1 day vs 3-4 weeks” comparison is a narrative estimate, not a repo figure, and is not asserted on these slides.

Primary contributors: Sam Schillace (both repos); Marc Goodner (~24% of the platform).