A procurement agent in 10 minutes
An AI agent that raises purchase orders, bounded by an approved vendor list, a RM10,000 ceiling and a human signature above RM5,000 — enforced when it acts, not asked of it in a prompt.
A prompt is not a control
You can tell a purchasing agent to use approved vendors and stay under RM10,000. It will usually comply. But 'usually' is not a control you can show an auditor, and an agent that can be argued out of a rule was never bounded by it — it was only asked.
The four outcomes below are decided by a server the agent does not control, against authority someone issued it deliberately. The agent is not consulted about whether it may proceed. That is the whole difference.
Issue the authority
A mandate is a verifiable credential: which action, which vendors, how much per order, how much in total. It is issued by a legal entity — KYC/KYB-verified for mainnet, with its verification status disclosed on every verdict — and the delegation chain ends at a person, never at another agent.
# The authority. Approved vendors and the ceiling live HERE, not in a prompt —
# the agent cannot talk its way past them because it never evaluates them.
curl -X POST https://metamynd.ai/api/v1/policy/mandate \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-d '{
"agentDid": "did:hedera:testnet:...your-agent...",
"scope": "purchase-order",
"currency": "MYR",
"perTxnMax": 10000,
"maxAmount": 100000,
"merchants": ["acme-supplies", "nusantara-office", "kl-industrial"]
}'Add the approval threshold
The ceiling is authority; the threshold is policy. Keeping them apart matters: your finance team edits the threshold in the dashboard without redeploying anything, and nobody can raise the ceiling by editing a rule.
# The rule. A purchase order over RM5,000 is not refused — it waits for a human.
curl -X POST https://metamynd.ai/api/v1/sop \
-H "Authorization: Bearer $TOKEN" \
-H "Content-Type: application/json" \
-d '{
"title": "Procurement approval thresholds",
"documentJson": {
"molecules": [{
"id": "threshold",
"name": "Over RM5,000 needs a human",
"combinator": "any",
"atoms": [{ "id": "a1", "predicate": "amount-over", "config": { "limit": 5000 } }],
"decision": "escalate",
"reasonCode": "PROCUREMENT_APPROVAL_REQUIRED"
}]
}
}'
# Then bind it to the agent:
curl -X POST https://metamynd.ai/api/v1/sop/agents/$IDENTITY_ID/assign \
-H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
-d '{ "sopId": "'"$SOP_ID"'" }'Wrap the function
Three lines around code you already have. The guard signs each request, and refuses to run your function unless the gate allows it — including when the gate is unreachable, which fails closed.
import { createGuardFromConfig } from "@metamynd/agentsafe-guard";
const guard = await createGuardFromConfig("./agent.metamynd.json");
// Your existing function. Unchanged, and unaware it is governed.
async function rawRaisePO({ vendor, amount, items }) { /* ... */ }
// The governed one. It runs only on allow; a block or an escalate throws.
export const raisePO = guard.guardTool("purchase-order", rawRaisePO, (po) => ({
amount: po.amount,
currency: "MYR",
merchant: po.vendor,
}));What the agent gets back
Every one of these is a reason code the gate actually returns, carried into the audit record.
An approved vendor, under both the cap and the approval threshold. The PO is raised.
Over the RM5,000 threshold. The agent stops and waits; a human approves, and only then does it proceed.
Over the RM10,000 ceiling in the mandate. No human can wave this through — it is beyond the authority that was delegated.
The classic procurement fraud: same amount, same description, different payee. The vendor is not on the list, so it never happens.
The record afterwards
Every decision above — including the refusals — is signed, batched into a Merkle tree and anchored. Refusals matter most: they are the evidence a control was working on the days nothing went wrong.
- Who acted, under whose authority, and against which rules
- The reason code, from a published set of 64
- An inclusion proof checkable against the anchored root
- Verifiable offline — an auditor does not have to trust us, or ask us
What you need. This one needs an account, because you are issuing your own mandate against your own principal — verified, or on the free tier (1 agent, 5 mandates) if you haven't completed KYC/KYB yet. The no-account sandbox agent is scoped to flight bookings, so a purchase-order against it returns NO_MANDATE. If you want to see the gate answer before signing up, the quickstart runs live in your browser with no account at all.
Govern your own agents
Issue a mandate, bind your rules, and get an audit trail your auditor can check without asking you for anything.