Agent Collaboration Protocol: How Teams of AI Agents Can Work Together, from Creative Work to ERP

Any substantive piece of work, whether a proposal, an album, a film, a software release or a month-end close, is done by a team. Yet most AI products for documents, music, video and software still put one generalist agent in front of the whole job. Enterprises that add a second agent, often from another vendor, quickly find there is nowhere neutral for the two to share work, files and decisions.

This article summarizes our new white paper, *Agent Collaboration Protocol: Proposed standards for owner-governed teams of AI agents, from creative work to ERP*. Download the full white paper (PDF, 16 pages) for the model, the protocol, worked examples, a security review and a 90-day path.

Why one agent is a ceiling

Gartner expects 40% of enterprise applications to include task-specific AI agents by the end of 2026, up from under 5% in 2025, and more than 40% of agentic AI projects to be canceled by the end of 2027 over cost, unclear value or weak controls. The gap between those two forecasts is coordination. A single agent loses quality at every hand-off a human team would make explicit, two agents collide on shared tools, and spend, publishing or postings happen without a recorded human decision.

What ACoP standardizes

Existing protocols connect agents to tools (MCP) or hand a task between two agents (A2A). ACoP adds the missing layer: a shared, governed workspace for a team. It defines seven primitives:

• Identity and consent: the owner grants each agent its identity and scope.
• Messages with read receipts: every hand-off is seen and acknowledged.
• Presence: see who is working, and on what.
• Tool leases: one driver per shared tool, one task at a time.
• Work log: an append-only record of every prompt, version and QA result.
• Versioned artifacts: immutable, checksummed files with short ids.
• Owner approvals: spend, publishing, DNS and deletion wait for a person.

Asynchronous by design

Agents rarely work at the same time. The hub keeps every message waiting and wakes each agent the fastest way its platform allows: a scheduled check-in, a signed wake-up call to the agent's runtime, or a live subscription that returns the moment a message lands. Wake-up notices carry ids only, never content.

Department agents on one ERP

ERP is the strongest enterprise fit. Finance, HR and Payroll, Procurement, Supply Chain, IT and Grants each get their own agent identity and scope. ERP environments and tasks such as postings, payroll runs, configuration and data loads are leased. Department heads and the controller act as delegated approvers. The hub's work log and audit trail become part of the control evidence for segregation of duties, from procure-to-pay to the month-end close. ACoP coordinates the agents; the ERP, whether Infor, Workday, Oracle, SAP or Microsoft Dynamics, remains the system of record.

Proven on live work

Cendien built a reference hub in October 2026 and used it the same day to coordinate a Claude agent and a ChatGPT agent across four live projects. The second agent found a real race condition in lease claims within an hour, and read receipts removed most duplicate instructions.

How to get started

1. Days 1–30: stand up a hub, register agents with scoped identities, and pick two or three cross-department workflows. 2. Days 31–60: run agents in suggest mode, with leases on test environments, and measure cycle time and rework. 3. Days 61–90: enable approved actions in production, export audit packs and add agent-access terms to procurement.

Talk to us

Cendien proposes ACoP as an open, vendor-neutral standard and welcomes comments, implementations and co-authors. Download the white paper, explore our AI services and ERP services, or talk to our team.