AI agents can increasingly discover, decide and transact. The missing infrastructure is a reliable way to establish whether they should.
mnd8t separates authority from access, so organisations can delegate safely without handing agents unrestricted control.
Express exactly what an agent is authorised to do.
Evaluate submitted payments and transfers against deterministic policy. Your protected executor must require a valid, claimed authorisation artifact before acting.
Maintain verifiable evidence connecting authority, decision and outcome.
Identity proves who an agent is. Payment infrastructure lets it transact. Neither decides what it is authorised to do.
Card networks, commerce protocols, agent identity standards and stablecoin rails are being built quickly and well. These systems carry or enforce parts of delegated authority. mnd8t's focus is a separate, versioned authority layer with stateful policy decisions and linked evidence — because whether this particular payment is within the authority a person granted depends on the budget already spent, this counterparty's history, and whether a human would have wanted to be asked.
That is the layer mnd8t is built for. We also maintain a public reference on the whole stack — the Agentic Authority Stack, a dated tracker, and analyses of AP2, ACP, Visa TAP and Mastercard Agent Pay.
Read the intelligence hubFive steps between an agent and the action it wants to take.
Delegate
A principal grants an agent specific authority in a mandate — approved counterparties, per-transaction limits, daily and monthly budgets, permitted purposes, operating hours. Versioned and immutable once published, and revocable at any time.
Request
The agent proposes an action and describes what it intends to do. Holding a payment method or an API key is not treated as permission.
Decide
mnd8t evaluates the request against the mandate and everything already spent under it. APPROVE, REJECT, or ESCALATE to a human — with the full reason set and a rule-by-rule trace. The model does not decide whether the action is acceptable. Policy does. Run it in shadow first; enforce when it has earned it.
Execute
An approved action receives a one-time signed authorisation artifact. Your executor verifies and claims the artifact, then executes with its own credentials. Today, supported execution is limited to a simulated wallet or EVM testnet; live-money support remains gated. mnd8t does not become the rail.
Prove
Every decision and every later state change produces a signed Authority Record, linking delegation, request, policy, decision and execution outcome.
Identity says who is acting. Payments say how. mnd8t decides whether it had the authority.
mnd8t occupies three stages of the chain — authority, the policy decision, and the evidence — and is built to sit across the systems that handle the rest. Change your model provider, agent framework, bank, wallet or payment rail without rebuilding the rules that define what your agents are authorised to do.
An approved payment is not a fulfilled mandate.
mnd8t doesn't stop at permission. Link what was authorised to reported execution, then evaluate the mandate's objective against customer-supplied evidence. mnd8t does not independently verify those reports.
An agent authorised to buy 100 units for no more than £10,000 can pay £9,700, on time, to the right supplier, entirely within its mandate — and report delivery of 80. Every authority system stops one step before that report. A mandate that records only what was permitted cannot tell you whether delegating it worked.
Today that means fulfilled or not, against evidence you supply. The Authority Record is the foundation for more: where execution systems can provide trusted evidence, the same chain can extend from authorised to executed, accepted, fulfilled, refunded or disputed. That is an extensible evidence model, not a claim that mnd8t independently verifies delivery.
Every decision leaves an Authority Record you can verify without us.
Not a transaction log. It records why the agent was permitted — or not permitted — to act: the principal, the agent, the mandate and its policy hash, the requested action, the decision and its reasons, any human approval, and the execution reference. Ed25519-signed and hash-chained; change one timestamp and verification fails.
Evidence, not surveillance. The record holds enough to answer what was authorised, what was requested, which policy applied, what was decided and what happened next — not everything an agent thought or did.
Designed to sit across these systems, not to become one of them.
That independence is strategic. The rules that define what your agents may do should outlive any one model provider, agent framework, bank, wallet or payment rail.
- an AI agent
- a bank account for agents
- a wallet
- a card network
- a payment processor
- a fraud-detection model
- an agent identity provider
- a merchant marketplace
Start in shadow. Enforce when it's earned.
A sandbox key, an agent, and a mandate. Most teams reach their first decision in under 30 minutes.
Get a sandbox key