AI / GitHub track · Microsoft Dublin · 10 Nov 2026

Agents need an identity, and your IT department will say no

Damian Flynn · Microsoft MVP, Azure · Innofactor

Damian Flynn

Damian Flynn

Practice Lead & Principal Consultant (Azure), Innofactor

30+ years in enterprise IT · cloud since 2009 · Mayo, Ireland
Azure Infrastructure as Code · Azure Networking · 17 years an MVP

Microsoft Most Valuable Professional
The request

I asked IT for an identity for my agent.
They said no.

No MFA, so it's a risk flag. Every request like it has been refused.
So my agent runs as me.

Your IT department isn't wrong to say no.

It's missing a way to say yes.

01 · Why agents run as you

Your agent is wearing your face

01 · Why agents run as you

What actually happens when your agent signs in

Agent on a laptopIDE agent, CLI harness, MCP client
token as
the human
Entra IDsees a normal user sign-in
audit
"Damian did it"via app X, if you capture the app ID at all

Nothing new was created. There is nothing to register.

Two populations, two control points

has its own principal

Managed / workload identity, Agent ID

Control the agent: registry, owner, scope, recertification, cost per agent.

borrows a human's token

Copilot, IDE agents, local harnesses

Control the client app: consent, conditional access, device, downscoped or on-behalf-of tokens. Log app ID + user.

Every over-permissioned account is now an automation platform

The agent's ceiling is the human's ceiling.
Your access debt now runs at machine speed.

02 · Why IT is right

What our CISO sees

Credentials with no human

An app identity can't do MFA, so it's a risk flag

Agents with no owner

The builder leaves; the agent keeps running

Agents nobody can name

He can count thousands; he can't name them

MCP is a black box

Weak logging, untested restores. Policy: "no MCP"

A "no" without a mechanism just moves the risk

Personal access tokens, API keys in .env files, agents on laptops.
The traffic still happens, just where nobody can see it.

03 · A way to say yes

One front door

Agentown Entra identity, or on behalf of a user
Entra
token
AI gatewayAPI Management: validate token, allow-list, token limit, log
managed
identity
Models + toolsFoundry · MCP servers · Jira / GitHub (PAT held by the gateway)

Back doors closed: Foundry keys off, agents get no direct RBAC. Entra in, PAT out.

Whose permissions does it have?

on behalf of

It can do no more than the caller

Owner = the caller. The safer default. If the delegation isn't granted: refuse, or ask out of band.

agent as employee

It has its own rights

Now you need RBAC for who may ask it to do what. Prompt injection is risk #1: a ticket says "delete the hub".

The back door you already have

One token. It can comment and close.

Whoever holds it can do both, and GitHub records that you did it.

GitHub fine-grained token: Issues read/write, Metadata read-only, Pull requests read/write, one repository
LIVE · sandbox tenant

Agents, one front door, five calls

1Through the gateway: agent, owner, tokens in one log line200 allowed 2Draft is allowed; "send" isn't in its contract403 denied 3Straight to the model, skipping the gateway401 refused 4Over its token budget429 limited 5Two agents, one GitHub PAT: one may comment, one may close201 · 403 · 200
Backup · recorded run

What the gateway logged

COMPLETED agent=agent-invoice-drafter owner=team-finance-ops op=chat-completions 200 tokens=46 AUTHORIZED agent=agent-invoice-drafter owner=team-finance-ops op=mail-draft DENIED agent=agent-invoice-drafter owner=team-finance-ops op=mail-send 403 operation-not-allowed DIRECT agent token → Foundry 401 · API key → Foundry 401 (keys disabled) REJECTED agent=agent-research-hog owner=team-research op=chat-completions 429 token limit DENIED agent=agent-support-triage owner=team-support op=github-close 403 credential=github-pat COMPLETED agent=agent-release-manager owner=team-release op=github-close 200 credential=github-pat
04 · Governance that holds

Three layers

1Org rulesetApplies to every agent: models, data zones, logging, owner required
2Agent contractPer agent, versioned: owner, identity, tools, data, budget
3RuntimeThe gateway stops it exceeding the contract, even when prompted to

"Draft a message" → "send a message" is a permission change, not a feature

Every change in authority gets a version, an approver and evidence.

From "no MCP" to a yes with conditions

before

"No MCP, please."

One approved exception. Everyone else waits, or goes round it.

after

"MCP only through the gateway."

With an owner, an identity, an allow-list and a log. The exception became the pattern.

Monday morning

  • List which apps can request tokens on behalf of your users
  • Close one PAT back door behind a gateway
  • Write a one-page contract for one agent

Your IT department isn't wrong to say no.
Give it a way to say yes.

damianflynn.com · github.com/damianflynn · linkedin.com/in/damianflynn