Problem
The agentic industry has advanced quickly in two layers:
- how an agent uses tools
- how multiple agents collaborate with each other
What remains poorly solved is something else: how an agent talks to people in real channels without breaking context, permissions, identity, tone, compliance, and traceability.
A bot in Slack doesn’t solve that. A WhatsApp connector doesn’t either. And far less five different bots that answer differently in each channel.
Novu named it this week: ACI, Agent Communication Infrastructure. It’s not a semantic curiosity. It’s an architectural signal.
Thesis
ACI matters because it turns agent‑human communication into its own layer of the stack.
MCP solves tool access. A2A solves collaboration between agents. ACI aims to solve the last operational meter: identity, threading, delivery, observability and governance when an agent must maintain a real conversation with people.
The consequence is clear: talking to users stops being an interface detail. It becomes infrastructure.
Framework
An ACI layer needs five components:
- Identity: know which person the agent is speaking with on each channel.
- Threading: maintain a single conversation even if the surface changes.
- Delivery: deliver on Slack, Teams, WhatsApp, Telegram or email without rewriting the agent.
- Governance: apply limits, compliance and observability.
- Escalation: decide when another human or another agent steps in.
Mini‑case: a user starts a conversation in Slack, continues by email and finishes on WhatsApp. Without a common layer, the agent treats each point as a new thread. It asks for context again, replies differently, or loses continuity. With an ACI layer, the agent inherits identity and conversation memory without being multiplied across five implementations.
Measurable signal: percentage of multi‑channel conversations that the agent maintains as a single verifiable thread.
Posture: the agent does not need more channels. It needs a layer that prevents it from becoming multiple broken bots.
Why it matters now
Novu launched Novu Connect on June 15, 2026 and positioned it explicitly as the first ACI block. The official thesis is straightforward: MCP connects agents to tools, A2A connects agents to other agents, and ACI connects agents to people.
That’s not just positioning. In its own material Novu emphasizes two concrete pillars:
- trust and governance for open‑ended messages
- communication experience that feels human and consistent across channels
If this layer matures, it changes the way we think about customer support, onboarding, internal sales, operations and software with customer‑facing agents.
Anti-example
“Connecting the agent to Slack is already done.”
No. The easy part is opening the channel. The hard part is everything underneath:
- who the user is on each surface
- what access they have
- whether the agent should continue, wait, or escalate
- how the same thread is joined across channels
- how an open and mutable conversation is audited
Without that, you don’t have communication infrastructure. You have sockets with personality.
Protocol (3 steps)
- Separate conversation channel. Slack, email and WhatsApp should not redefine the agent’s logic.
- Design identity and continuity. The same user must be recognizable across surfaces.
- Instrument governance from the start. Logs, permissions, escalation and policies cannot be post‑production.
| Layer | Question | Risk if missing |
|---|---|---|
| identity | who the agent talks to | confusing users and permissions |
| threading | which thread stays alive | repeat context or fragment it |
| delivery | where it replies | duplicate implementations |
| governance | what limits it has | unsafe messages or non‑compliance |
| escalation | when another actor enters | useless conversation loops |
Related
- A2A + MCP: agent protocols are not a product, they are the nervous system
- Microsoft Foundry Local + Scout: when agent work moves to the perimeter
- Tool Registry: the new risk map for enterprise agents
Sources consulted
Next step
Do an uncomfortable test: follow an agent conversation with the same person across two different channels. If the system cannot maintain identity, context and limits, you still don’t have agentic communication. You have messages.
Translated from the Spanish original with AI assistance and reviewed for accuracy. Read the original in Spanish.