Risk control for agentic AI

Sheaf is the Coherence Layer for Agentic AI.

Your AI agents can contradict each other and still sound right. Sheaf measures that, and stops the action before it happens.

Your agents work on the same decision. Sheaf measures where they agree and where they contradict each other. When a contradiction matters, Sheaf stops the action before it happens and hands a person the reason, with a signed record of what was stopped and why. Agents that contradict each other are a risk. Sheaf measures that risk and gives you a way to stop it.

Check how safely your agents can act →

Seven questions, two minutes, no account. You get a score, the weaknesses it found, and what Sheaf would do about each one.

Three ways to connect Sheaf → Measure · Gate · Orchestrate

Why we built Sheaf · a 46-second film

This is the moment Sheaf is for: a bad decision is stopped before the money moves. Watch it stop an action, live →

The problem

Agents you cannot stop, and a record you cannot defend

Businesses are moving AI agents from trials into daily use. These agents do more than answer questions. They take actions with real consequences. Once you allow that, two new questions matter more than “was it a good answer?”

Today the answer to both is no. You get agents that act, and you get a transcript of what they said. The transcript cannot stop anything. It also cannot show that two agents drifted into different pictures of the facts, each sensible on its own, and wrong when put together. This is where agentic AI stalls in a regulated business.

What Sheaf is

Sheaf checks whether the agents' reasoning holds together.

Sheaf watches over a group of working agents and does two things. It stops an action before it happens when a person's approval is required. And it measures whether the agents were all working from the same set of facts, as a score you can put on a page.

Many products call themselves “the trust layer for AI.” Most of them check identity: is this agent who it claims to be, is it allowed to act, can we prove it acted. That is useful work. Sheaf checks something else: the agents' coherence, meaning whether what they say fits together. That is what decides whether you can trust the result.

How it works

Five parts around every agent

The stop is instant and simple, so it is reliable. The checking sits beside the work, so it never slows the agents down.

Ledger
Before anything runs, you list the facts the agents work with, such as an applicant's income, the loan-to-value, the rate and the risk tier, and the actions that need a person's approval. Everything Sheaf measures is measured against this list.
Gate In front of the action
Sits in front of every action. It is simple and instant. If an agent is about to take an action on the approval list, the Gate stops it before it happens. No AI model is involved, so there is no delay and no judgment call to get wrong.
Witness Beside each agent
One for each agent. It reads what its agent is thinking and doing, and writes down that agent's claims about the listed facts, meaning what the agent believes to be true.
Warden Calls a person
When the Gate stops an action, or the agents contradict each other, the Warden pauses the work and puts it in front of a person, who can approve it, refuse it, or step in. No important action happens without that step.
Reporter Checks all the time
Collects what every Witness has written down and keeps checking that it fits together. If two agents have drifted apart, it catches the contradiction while it is happening and calls the Warden. When the work ends, it issues a signed record of whether the reasoning held together.
Why this works in practice: the stop is simple and instant, so it is reliable. The checking runs beside the agents, so it never slows them down.

The record

A score a risk officer can sign

When a run finishes, Sheaf issues a signed record built on two measurements. A model cannot make them up about itself.

Consistent

Every agent's claims fit together as one picture (the mathematics calls this H⁰).

1.00coherence score: everything fits
Contradiction

Any two agents agree, yet no single picture holds all of them (the mathematics calls this H¹).

0.41coherence score: a contradiction was found

The mathematics behind it is also used in engineering to combine readings from many sensors and to check that designs contain no contradictions. Sheaf applies it to agents that act. That is what lets a risk officer sign the record, a regulator read it, and a board rely on it.

How you connect it

Three ways to put Sheaf in front of your agents

First decide who runs the agents. If you keep your own agents, Sheaf watches from outside. It can warn you, or it can stop the action. If you let Sheaf run several AI models for you, the stop is built in. All three return the signed record.

Measure Your agents · warns you
Send your agents' outputs to one Sheaf address. Sheaf returns the signed record and points out contradictions. Your system decides what to do. Nothing changes about how your agents run. The limit: if the agent acts anyway, Sheaf has recorded the warning but has not stopped the action.
Gate Your agents · stops the action
Sheaf sits between your agent and its tools and stops a contradictory action until a person releases it. The agent cannot skip it. It connects through MCP, the standard way tools are connected to an AI agent, and setup takes minutes.
Orchestrate Sheaf's models · stops the action
Sheaf runs several AI models for you and returns one answer. Tell Sheaf which action the answer leads to, and a contradiction stops that action on its own. It needs no change to your systems. The measurement, the answer and the stop come from a single request.
All three measure. Keep your agents: Measure warns you, Gate stops the action. Let Sheaf run the models: Orchestrate stops the action.

See it work

Watch the measurement catch a contradiction.

Ask any question. Five leading AI models answer it separately, and Sheaf measures how well their answers fit together. It calls this the H¹ inconsistency index. It is the same measurement Sheaf runs over your own agents. It shows you where the answers agree, and the exact contradictions that stay hidden when you only compare two answers at a time.

The live demo

A live measurement across five leading AI models, in about a minute. No sign-up.

Try the live demo → Public · 5 models · live measurement

Prepared demos for regulated work, such as AI mortgage advice and AI-run renewals, shown on request.

What Sheaf checks

Three checks on every decision

Sheaf asks three questions about every decision your agents make. Does it match your source of truth? Point Sheaf at the facts you treat as correct, and it flags any agent whose claims contradict them. Is it possible? Even with no source, it flags values that are clearly false, such as a £1,000-trillion cost or a 400-year-old applicant. Does it hold together? It measures whether the agents were all working from the same set of facts, and returns one signed record of the result.

This catches the failure that causes incidents: agents that each look fine on their own but no longer agree with each other underneath. It runs on every case, at machine speed, and records that it checked.

“We use AI agents” and “we can show our agents are safe to use” are two different statements.
Sheaf gets you from the first to the second.

Sheaf, the coherence layer for agentic AI.
It stops the actions that need a person's approval. It records whether the reasoning held together.