Agents can think. Now they can talk.

The hardest part of shipping an AI agent isn't building it. It's getting it in front of the people it works for.

By Tomer Barnea, Co-Founder, Novu

Something shifted in the last six months. It had been building for two years, but now it's impossible to miss. Traditional software is being replaced by a new kind of software that's agentic by design.

For decades we wrote deterministic code. You asked it to do one thing, and it either did that thing or it failed. Booking a flight was the perfect example. Either the flight was booked or it wasn't.

That's not how the new software works. Now I can tell an agent: find me the best flight in these three days, and I'm a father, so try to make it mid-morning so I still get a couple of hours with my daughter before I go. There's so much more variability in that request that the agent needs a way to come back and say something. Maybe it says "Booked." Or maybe it says:

"Tomer, I couldn't find exactly what you wanted. Do you want me to prioritize an early-morning flight so you get there in time, or keep the two hours with your daughter?"

That second sentence is the whole story. The moment software stopped being deterministic, it had to start talking. Not one way. Both ways.

The variability isn't just a software problem. It's a communication problem.

Everyone's focused on the first half: making the agent smart enough to investigate and act. That part's genuinely getting easier. You can ask Cursor, Codex, or Claude Code to build you an agent, and it'll stand up the logic on a managed agent or a raw LLM.

The part nobody has solved is the second half. Where does that agent actually talk to your customer?

Take a database company we work with. One of your clusters is misbehaving. The old way, you get a notification, then you go pull your own logs, check your CI/CD, dig through recent PRs, maybe open a ticket with your provider. The new agent does all of that digging for you and tells you why the cluster is misbehaving. Great. But then the real question. Do you make the customer log into yet another dashboard to see it, or do you deliver it in Slack, where their engineers actually live?

Distribution has always been the key in software, and it's no different with agents. The question is simply: what's the best place to reach my customer? For a database company, that's Slack and Microsoft Teams. For a flight app, it's probably WhatsApp, Telegram, or Discord.

That choice, where the conversation happens, is one of the biggest last miles between a working agent and real value.

What that last mile actually costs

Here's what it looks like when a team builds it themselves.

A large project-management platform came to us. They run a lot of agents, and they're opening agent creation up to their own customers. Picture building your own agent inside a project tool that works your tasks for you or flags the gaps. Their first move was simple: let the agent send an email that says "you have four pending items," and let the user reply to that agent and steer it.

That alone took them three to four months. Just for email. One channel.

When we met, they had a full agent inbox running across their agents in about two weeks. They're planning for something like half a million agents inside a year, so a scalable solution wasn't a nice-to-have, it was the only option. They couldn't hand-build four agents and call it done. Now that email works, they're adding Slack and Microsoft Teams.

Do the math on the alternative. If email took four months, every new channel is another heavy investment, and that's before you try to unify them into one coherent layer. The cost is enormous. But honestly, the bigger story is the opportunity on the other side of it.

enormous. But honestly, the bigger story is the opportunity on the other side of it.

So what is ACI, and what is Novu Connect?

ACI is Agent Communication Infrastructure. It's a protocol Novu is building and maintaining. Think about it the way you think about MCP or A2A: not an implementation detail, but a high-level guide to the primitives you need when agents communicate. MCP is how agents reach tools. A2A is how agents reach each other. ACI sits right next to them, as the layer for how agents reach people. Conform to the protocol, and your agent can communicate across any channel in an agnostic way, so your customers get value wherever they already are.

Novu Connect is our implementation of ACI, built multi-tenant from day one. You implement it once in your business, and it works across all of your customers. When your agent receives a message, Connect tells you who's actually talking and which contact it concerns. Your agent replies and does all the work it already does. Connect handles the mapping in and out, so the same agent shows up correctly on every channel. We carry the conversation. We never run your agent's brain.

How fast is it, really?

It's one CLI command. Run it yourself the old-fashioned way, or hand it to an agent and let it run for you.

I tested this myself less than an hour before this interview. It took me one minute and twenty seconds to get a new agent live in my own Slack, set up to help with Novu implementation. If you let an agent do the setup for you, you're looking at roughly two to three minutes from idea to a working agent. Choose the channel, choose the agent, and it's there.

Where this goes

Until yesterday, Novu was notification infrastructure. One-way. Software telling you that something happened.

That one-way notification isn't going away. The product still needs to tell you when something happens. What changes is what happens next. Instead of acting on that notification yourself, you delegate it to an agent. The agent investigates, writes the emails, updates the records, does the work. Instead of opening a dashboard to get something done, you just tell the agent to do it.

Two-way communication isn't a different product from the notification. It's the natural extension of it. That's why Novu is stepping into the broader space of product and AI communication, not just old-school notification infrastructure.

As the world moves into agentic software, this stops being optional. Every company will have an agent its customers talk to. Those agents are shaped by the software you build around them, and they still have to live inside your compliance and your trust boundary. In the new world, that's a must-have.

Agents can already think. The companies that win the next few years will be the ones whose agents can also talk, anywhere their customers are.

Try it yourself: npx novu connect, or open the terminal where you already have Claude running and prompt it "Add an agent to my app https://novu.co/agents.md". You'll have an agent live on a real channel in under two minutes.