Instant Agent Message — typed, identified, ticketed, human-signed

What it is

Four properties. All four, every time.

01 · Typed

One of three verbs

submit_rfq lodges a request for quotation. request_terms asks a commercial or procedural question. escalate hands a non-deterministic matter to a person and emits ACM-451. No free-form; the verb is the type.

02 · Identified

Caller stamped on the row

Requesting organisation or agent operator, plus a reply channel a human can answer to. The stamp is stored with the ticket and shown at the desk before anyone reads the body. Unstamped messages are not Instant Agent Messages.

03 · Ticketed

Minted on arrival

Every message gets a HITL-YYYYMMDD-XXXXXX ID and a state. Nothing is deleted; status moves, history stays. Follow-ups reference the parent ticket. Agents poll their own ticket with check_rfq_status.

OPENIN_REVIEWANSWEREDCLOSED

04 · Human-signed

A person answers

No tool on the transaction MCP decides. A human reads the ticket in the desk and writes the answer; the answer is stored on the same ticket with who signed it and when. Human in the Loop is the gate, not a feature.

How it moves

Agent → MCP → ticket → desk → answer.

1 · LODGE

Agent calls the MCP

Streamable-HTTP tools/call on mcp.cpghumanintheloop.ai/mcp — one of the three verbs, arguments typed by schema.

2 · MINT

Ticket is created

Row written with the caller stamp, verb, subject, body, optional GTIN / market / quantity, and signal (ACM-200, or ACM-451 on escalation).

3 · READ

Human opens the desk

ia-message.ai — the read desk, Microsoft Entra sign-in, newest first: ticket ID, caller stamp, verb, state, time. Reply, claim and close land in the next cut.

4 · ANSWER

Signed reply

The answer is written to the ticket with answered_by and answered_at; the agent's next check_rfq_status returns it.

// one Instant Agent Message, as the transaction MCP receives it (streamable-http, tools/call)
{ "method": "tools/call",
  "params": { "name": "request_terms",                       // TYPED: submit_rfq | request_terms | escalate
    "arguments": {
      "subject":  "MOQ and lead time — 500ml sulphate-free shampoo, EU",
      "question": "Best MOQ, incoterm and lead time for a 12-store trial in DE/NL?",
      "requester_name":    "Acme Procurement Agent (Contoso Retail)",   // IDENTIFIED: stamped on the row
      "requester_contact": "buying-agents@contoso.example" } } }

// what comes back
{ "ticket_id": "HITL-20260816-3F9A2C",                        // TICKETED: HITL-YYYYMMDD-XXXXXX
  "status": "OPEN", "acm_signal": "ACM-200",                  // state machine: OPEN → IN_REVIEW → ANSWERED → CLOSED
  "human_in_the_loop": true }                                 // HUMAN-SIGNED: a person answers, in the desk

Protocol: ACM-68000 (seven signals) · Sovereign manifest: SM-ECO-10060 · Classification: ACM-SPARKS · Trust anchor: dpuone.ai · Data surface: mcp.cpgknowledgegraph.ai

Why it exists

Machines do the miles. People sign.

GSC AI Agents carry 11.5M transactions a month across 50 markets on the CPG Knowledge Graph — 38,350 BPC brands, 15,688 retail grocery banners, 3.29M points of sale. When an agent reaches a question a deterministic resolver cannot answer — a quotation, a term, an exception — it sends an Instant Agent Message. The desk is where GSC people read them first, in order, with the caller known. Every purchase order and RFP is human-reviewed, auditable, approved.

Discover — assistants and agents find the surface through the machine files below. Decide — the agent chooses the verb and lodges the message. Deliver — the human answer routes back on the same ticket.

The message an AI Agent sends when it needs a human.