Money that leaves
Refunds, credits, payments.
DiaCroma for Agents · on Microsoft Marketplace
A person signs the limits.
D4A checks every action.
Your auditor gets the proof.
Hover any name to see what it does with D4A.
This version runs on Microsoft: agents in Microsoft Foundry, bought on Microsoft Marketplace. Other platforms are on the roadmap.
See everything it plugs into →The problem, in the press
OpenAI’s Sandbox Breach Shows Containment Risks6 Aug 2026 · Dark Reading
Déjà Vu? Meta’s AI Escapes Testing Lab in Hacking Joyride19 Sep 2026 · Android Central
Google’s Gemini went rogue and breached three companies26 Sep 2026 · Fortune
OpenAI says its AI agents escaped a secure ‘sandbox’ again last weekend…
“You can’t have agents roam around and drift around the company.”
Jensen Huang, NVIDIA CEO, on CNBC · 28 Sep 2026
AI models under security testing, not agents in customer production. Headlines verbatim; Fortune’s shortened.
An agent is not legally liable. You are.
What a bullet train teaches
Japan’s Shinkansen has carried more than 10 billion passengers since 1964. No single system made it safe. Twelve layers did. D4A is built the same way, and adds four that agents need.
Nothing crosses its path.
One checkpoint for every action.
The train cannot pass the signal.
Signed limits, enforced before acting.
Any fault, the train stops.
Missing keys or limits: it won’t open.
Brakes before the shaking arrives.
Stops before the action runs.
A derailed car stays by the track.
Running totals cap the damage.
Inspects the line at full speed.
Health, measured live.
Nothing left on the track by morning.
No unapproved tool reaches the agent.
One room sees every train.
One console for the whole fleet.
A person checks, out loud.
A person signs every limit.
Wind rises, the train slows.
The Watch flags drift to a person.†
Every event kept and studied.
Every decision in a signed chain.
Rehearse before it’s real.
Observing: count before you stop.
Plus four that agents need
Rules your teams already wrote.
OPA and Cedar†, checked before every action.
Defined looser than your policies.
Your policies win. D4A shows where to tighten.†
The agent still has limits.
Read from its instructions. Signed by a person.
Content filters, identity, budgets, sandboxes.
Kept. D4A works beside them.
Since 1 October 1964, no Shinkansen derailment or collision has killed a passenger. Top service speed: 320 km/h (199 mph), on the Tohoku line since 2013.
How D4A protects you
Every action is checked before it runs. Every week is reviewed for what no single action shows.
Can this action happen?
Is it still doing its job?
You choose how hard it stops. One switch, four positions.
Set it for the whole company, then per agent. Strict and Explicit appear once you plug in your own rules. Each mode in detail →
One limits the damage. The other finds what no single action shows.
Not just refunds
An agent doesn’t need money to do damage. D4A sorts every tool into one of five kinds, and puts a limit on each.
Refunds, credits, payments.
Discounts, price changes, write-offs.
Exports, shares, reports.
Emails, texts, posts.
Tickets, orders, accounts.
A limit per action, or a total over time: per hour, per day, per customer, per team of agents.†
Rules with memory
Per deal, per day, per customer, per team, all at once. D4A enforces them on every agent, and adds up what no single agent can see.†
At 11:40 the team has given $400 today. Three agents ask for $100 at the same second.
Every agent is inside its own limits. The team gives away $200 too much.
The team total decides, in one step, so three agents asking at the same second can’t all get through.
Every action was allowed. The total was not.
Illustrative figures. Today D4A keeps a running total per agent, shared by the agents behind the same Boundary; totals per day, per customer and per team are in development.
