SHADOW

Inspiration

AI agents can already click buttons and change data. What bothered me was the step before that: how can I review the outcome before it becomes real?

In software, I would create a branch, inspect the diff, and merge it after review. SHADOW brings that workflow to AI actions in the browser.

What it does

SHADOW gives an agent an isolated copy of the current state. The agent can change plans, reduce seats, or cancel subscriptions without touching Reality.

In the demo, the fictional company ORBIT spends $184,300 per month on software. The agent reads usage, contracts, and dependencies, then builds different cost-saving proposals. Every change includes Proof showing the data behind it, its impact, and any blockers.

Reality → Shadow → Proof → Human commit

Only the user can commit a Shadow. The agent has no commit tool.

How I built it

I built SHADOW with Next.js, React, TypeScript, Tailwind CSS, Zod, and WebMCP.

A deterministic projection engine calculates each future without changing the original state. I exposed 13 WebMCP tools for reading data, staging changes, creating branches, and comparing outcomes. The UI and the tools use the same application logic, so they cannot produce different results.

The MVP runs locally in the browser with synthetic data and needs no accounts or API keys.

Challenges I ran into

The hardest part was making the separation between Shadow and Reality real, not just visual. I had to ensure that no agent-accessible path could commit a change.

Unsafe proposals were another challenge. I did not want to hide them because they often contain useful information. Cancelling Miro, for example, reveals a $25,000 penalty. SHADOW keeps that proposal visible but blocks it from being committed.

I also had to make projections reproducible and mark old Shadows as stale whenever Reality changes.

Accomplishments that I’m proud of

I am proud of what SHADOW deliberately cannot do. Its 13 tools let an agent inspect, stage, branch, and compare, but none can alter Reality.

The full flow works end to end: an agent creates two alternatives, the user combines the best changes into a Hybrid, and a human commit produces a new Reality version and receipt. The Hybrid reduces monthly spending from $184,300 to $147,385.

What I learned

I learned that human oversight works best when the system enforces it. A prompt alone is not a safety boundary.

I also learned that useful Proof is not hidden model reasoning. It is the visible data, impact, and checks that let someone make an informed decision.