Vibe Code, Safely

Governance as the moat for enterprise AI coding

The Frame

Your people are already vibe coding

AI now turns "describe an app" into a working app. Employees will build this way whether IT sanctions it or not — the only real question is whether it happens under governance or in the shadows.

So what happens when it happens in the shadows?

The Tension

Ungoverned, that work lands in the shadows

Without a sanctioned path, vibe coding runs on shadow AI and personal accounts — and the enterprise loses the three things it depends on.

The response has to restore exactly these. It already exists.

No identity — work runs on personal accounts, unattributable
No audit — nothing recorded, nothing reviewable
No compliance — no scanning, no policy, no controls

The Proof

The answer already ships as working code

This is not a proposal. A first wave of 16 module packages exists in src/, with roughly 13.6K lines of non-test Python behind it.

Those 16 modules divide cleanly into two halves.

16

module packages (9 hooks, 6 tools, 1 orchestrator)

13,584

lines of non-test Python source

6,008

lines of test code across 42 test files

The Platform

The Code-Safe Platform splits the world in two

The Amplifier M365 Code-Safe Platform hands coders safe, isolated sandboxes — and hands platform teams the validation and governance layer. Two sides, one discipline.

Start with the coder side: how isolation is made real.

Coders — safe, isolated sandboxes to build and run in
Platform teams — validation and governance over every action
3 agents ship — vibe-coder, app-reviewer, rollback-specialist

Mechanism — Coder Side

Builders get a place code can't escape

The coder side is real isolation: a sandbox execution tool runs the code, and a tool-permission gateway walls off each workspace so builders can experiment without reaching production.

Now the other half — governance for the platform team.

Sandbox tool — a code execution engine for sandboxes
Tool-permission gateway — workspace sandboxing walls off each build
Shadow environments — OS-level isolation from production

Mechanism — Platform Side

Every action becomes attributable and reviewable

The platform side restores exactly what shadow AI stripped away: Entra identity, RBAC, compliance scanning, and audit turn each action into an event with a name, a permission, a policy check, and a record.

Real code, but does it actually work? The evidence follows.

Entra identity + RBAC — attribution and authorization
Compliance scanner — policy enforced as a hook
Audit + audit-export — every action recorded and reviewable

The Payoff

Governance is the moat — and it's demonstrably real

Not a deck of intentions. The modules pass their unit tests, the protocol adapter validates fully, and E2E runs against a real Microsoft 365 tenant — all built across 37 commits in about 8–9 days.

Which points at the durable lesson.

227

unit tests passing (3 orchestrator tests failing)

18/18

protocol adapter + orchestrator tests pass

4/5

simplified E2E PASS + 1 PARTIAL on a live tenant

~8–9

days to build the first wave (2026-01-15 to 2026-01-23)

The Takeaway

Governance, not generation, is the moat

In the AI age everyone can generate code. The durable advantage belongs to the enterprise that can make that code safe, compliant, and governed — and a working M365 platform shows it can ship.

Vibe coding is coming either way. Meet it with a discipline.

Sources

Sources & Research Methodology

Status: First wave shipped & locally validated

Data as of: 2026-01-23 (latest commit). Source repo: ramparte/amplifier-m365-platform ("Amplifier M365 Code-Safe Platform").

Research performed (commands run):

Module inventory: ls -d src/amplifier_* | wc -l → 16 (9 hooks, 6 tools, 1 orchestrator)
Bundle wiring: grep -E 'module:' bundle.md ; grep -i 'orchestrator' bundle.md → 15 M365 modules wired; orchestrator "available but not yet integrated"
Source lines: find src -name '*.py' -not -path '*/tests/*' | xargs wc -l | tail -1 → 13,584
Test lines: find src -name '*.py' -path '*/tests/*' | wc -l → 42 files, 6,008 lines
Unit tests: uv run --with pytest ... python3 -m pytest src/ → 227 passed, 3 failed
Validation: cat VALIDATION_RESULTS.md ; head -60 FINAL_E2E_EVIDENCE_REPORT.md → 18/18 adapter+orchestrator; E2E 4/5 PASS +1 PARTIAL (real tenant 1627d542...)
History: git rev-list --count HEAD → 37 commits; first 2026-01-15, last 2026-01-23
Contributors: git shortlog -sne --all → Sam Schillace (28), Marc Goodner (9, Microsoft)

Gaps & caveats: 3 orchestrator unit tests fail (async/timestamp mocks); SharePoint E2E is PARTIAL; 5 modules needing cloud SDK deps weren't run in the minimal-deps pass. "VibOps" and "first wave" are narrative framing, not codebase terms.

Primary contributors: Sam Schillace (28 commits) with Marc Goodner / Microsoft (9 commits).