The missing layer between agents and people

I remember the night this started.

A cron job was running. One of my co-founders raised the alarm in the middle of the night. We'd accidentally sent hundreds of thousands of notifications to people who should never have received them.

I lay there thinking, there has to be a better way. We'd reinvented the wheel of notifications so many times by then. That night is where Novu started. One platform that bakes in the best practices and guardrails for how products talk to people.

The assistant we couldn't build

Early on, we ran a whiteboard session and asked ourselves a simple question. If there were no limits, what would the best notification platform look like?

The answer we kept coming back to was a great personal assistant. Think about how a good assistant behaves. If you're in a meeting and a call comes in, she doesn't burst into the room and say you have to answer right now. She waits until you're done. She aggregates the things that need your attention, brings them to you at the right moment, and lets you answer back. This one isn't relevant. Remind me later. Get me more context on that one.

Five years ago, that was the vision. The problem was a technology gap. We didn't have a way to actually build it, so we built the closest things the technology allowed. Digest. User preferences. Time-zone awareness. Useful, but never the magical assistant moment.

The reason it stayed out of reach is that notifications were always a guessing game. There was a product manager, the mastermind, trying to decide the perfect content and the perfect moment to send. But the right content for a notification is different for you and for me, based on who we are and what our role is. It was a game you could never win.

When the technology finally caught up

For me, the moment it became undeniable was about trust.

Over the last two and a half years, there was a slow transition of trust in the models. I remember the first GPT models. I couldn't trust the output, and more importantly I couldn't trust them with my instructions to act on my behalf. Then, around the start of this year, with Claude Opus 4.5, I felt it for the first time. I could trust it to do things for me.

That was the moment, internally, when we realized the technology gap we'd described five years earlier had finally closed.

The shift we're seeing across the industry is the shift from notification to communication. Software is moving from mouse clicks and dashboards to something you can actually talk to. And the infrastructure for that doesn't exist yet.

From notification to communication

A notification used to be the only way a product could reach you. To me, a notification is a necessary evil. It's a technical concept. It isn't why engagement goes up or why people succeed with your product.

The interesting part is what happens next. You get a message that a new lead came in, and it feels completely natural to reply. Assign this to her. Give me more context. Go check that one. You want to have a conversation about the notification. That conversation is where the value is.

So what is ACI?

When we built open-source notification infrastructure, the gaps were easy to name. Send a message. Manage the content. Add rules for digest. We built those primitives.

Agents reopened every gap. The thing producing the message isn't an API sending a structured event anymore. It's an entity whose output you can almost never predict. The content is unstructured. The conversation can go on and on. There's no product manager defining the exact words and variables.

So the questions become. How do we bring structure back to this? How do we build trust around it? ACI, Agent Communication Infrastructure, is how we think about that, and the tools we give companies so agent communication doesn't turn into chaos.

It has two pillars.

The first is trust and governance. Notifications were rigid, which made them safe. You could do compliance. You could make sure they didn't leak data. We need to bring that same trust to agent communication, where the output is open-ended.

The second pillar matters even more to me. The communication experience itself.

Think about how you talk to your teammates on Slack. You add an eyes emoji. You drop an hourglass to say you're working on it. You use memes, reactions, links. That's how humans express emotion. How do we bring that to agents?

We're a fully remote company, so we live and breathe async communication. We've spent years learning what good communication between people actually looks like, and we even consult behavioral psychologists on it. ACI is us taking those human-to-human lessons and applying them to agents talking to people.

MCP, A2A, and the missing layer

MCP connects agents to tools. A2A connects agents to each other. ACI is agents talking to people.

A lot of the industry is focused on the technical plumbing of connecting agents to humans, and there are incredible startups solving identity and security and connectivity. Our business is the most basic primitive underneath all of it. People actually speaking with agents, over the channels where those people already are. That's the agent-to-human layer, and it's the one nobody owns yet.

Novu Connect is our implementation of it. We carry the conversation. We never run your agent's brain.

Connecting to Slack is the boring part

We've been sending messages to Slack for almost five years. That's not the exciting part. The exciting part is everything underneath the words "just connect my agent to Slack."

Making it feel natural on Slack, WhatsApp, Telegram, and iMessage without you building any of it: reasoning cards, typing indicators, editing a message, reacting with an emoji. Working out who you're talking to right now, what access they have, and whether you need a human in the loop or a third party pulled in over another channel. Pulling several channels into one conversation, sometimes with multiple participants and even multiple agents in the same thread. And knowing what's happening across all of it, keeping the conversations safe and checking they're doing what they should.

We're in the early days of understanding how deep this goes. Every week the field changes, with new patterns and new APIs. Our team wakes up in the morning and the only thing we care about is making agent communication better. That's the point of a partner like Novu. You focus on your edge, you leave the communication to us, and one morning you wake up to a better system for your users without doing the work yourself.

What people are building

I think about agent maturity in levels. Level one is an internal agent for your own team. We already have twenty to thirty of them in our Slack. Level two is a customer-facing agent that helps your users get something done inside your product. Level two and a half is connecting that agent to the tools and MCP servers your customers already use. Level three, the most advanced, is letting your end users create their own task-specific agents inside your product. Novu gives you the low-level primitives for all of it.

Why open source, and where it goes

If you asked me why we built notification infrastructure in the open five years ago, I'd give you the same answer I give today. We want to learn from the world. Building a real standard takes the input of millions of developers telling us their challenges and their use cases. That's why we started Novu in the open, and it's exactly why we're building ACI in the open now.

Three years out, the core stays the same, just amplified. Agents will be everywhere. In your fridge. In your car. In the humanoids walking through the room. The question for all of them is the same one we're answering today. How do we communicate with them?

Try it in your terminal. npx novu connect. Or just tell your favorite agent, "Add an agent to my app https://novu.co/agents.md." That's it.