For thirty years a company's presence on the internet pointed one way: inward. You built a site so people would come to you, read it, and buy something. Everything about the web, the markup, the design, the analytics, assumed a human turning up.

That quietly stopped being the whole picture. A business today still takes in hundreds of thousands of customers through the front. It has also started sending agents out the back, into the same internet, to go and do things for it.

Zatanna sells the equipment for the second direction.

What they built

The method is simpler than it sounds. A person completes a task once inside some ancient supplier portal or carrier system, while Zatanna records what the browser is really doing underneath. The company then turns that traffic into a clean, stable connection, handling logins, sessions, two-factor codes and retries. After that, an agent makes one call instead of driving a browser around.

When the portal changes, the system spots the break and repairs it. Requests come back in milliseconds rather than the seconds a browser needs. The customers are companies automating back-office portals, business systems, freight carrier sites and insurance platforms, described by Zatanna as fast-growing startups in healthcare, logistics and hospitality.

The number they publish is 5,236,271,750 requests since the first of March. Whatever weight you give it, that is not the shape of a pilot.

Who else is in this field

Connecting software to software is a crowded business. Zapier made it something anybody could do. Merge, Nango, Paragon and Unified.to sell ready-made connections to developers. On driving a browser rather than calling an interface, Browserbase, Browser Use and Apify run the machinery, with Bright Data supplying much of the network underneath. Anon works on letting an agent log in as you.

Most of them connect to systems that wanted to be connected to. The awkward half is the portal that never offered a way in.

The point: the web now runs both ways at once

The interesting claim is not about the quality of the connections. It is about what the internet is now for.

Because of AI, the web has stopped being a place people read and become a surface that software operates. A business no longer just publishes and waits. It sends things out. Buying agents checking supplier stock, operations agents filing with carriers, finance agents pulling statements, support agents settling claims across three systems that have never talked to each other.

The same internet a company built to invite customers in is now the internet it sends agents out through, and both run at full volume at the same time.

Almost every tool that exists was built for the inward direction. Analytics count arrivals. Content systems format things for readers. Security assumes a person or an attacker, not a fleet of approved agents acting for a customer. The outward direction has almost no infrastructure at all, which is odd given how much money is going into the agents themselves.

Zatanna is one of the first companies building for it, and the choice of layer is the giveaway. Driving a browser is what you do if you think of the web as something to be looked at. Capturing the traffic underneath is what you do if you think of it as something to be operated.

Why the idea is easy and the business is not

It should be said plainly that this is not a hard idea to have. Unofficial workarounds for closed portals have existed for as long as closed portals have. Anyone technical who has been forced to automate a supplier system has built one.

They have also, every single time, rotted. The portal changes its layout or its login flow, the workaround breaks quietly, and somebody loses a Thursday afternoon working out why the numbers stopped arriving. That is the entire history of this idea, repeated in thousands of companies, and it is why nobody treats it as infrastructure.

So the hard part was never working out how the portal behaves. It is the upkeep, and the upkeep is the actual product.

That changes how to read the company's boldest claim, which is that it works against any portal, including ones it has never seen. As a technical boast it invites doubt. Read properly, it is a description of a service. You bring them the system you are stuck with, they capture what your own session already produces, and they carry the repair work from then on. The customer is not buying a connector. They are buying the removal of an ongoing headache, which is a much easier thing to keep paying for.

The open water behind it

A prediction now, offered as a prediction and not as reporting.

If the outward direction is real, then nearly every piece of infrastructure the inward web has will eventually need an outward twin, and none of them exist yet. Identity and permissions for agents acting for a company. Rate limits and good manners between a business and the systems its agents touch. Monitoring for outward work. Payments between machines. Records of what your agents did while you were asleep.

Expect that to reach where the agents run, too. A lot of this work will end up on a company's own machines and servers rather than in a vendor's cloud, for reasons of cost, data control and plain nervousness about handing a third party the keys to every portal you use. Serving agents that run locally is a different product from serving a hosted platform, and almost nobody is building for it.

That is a genuinely open field. The inward web took three decades to build up its tooling. The outward web has almost none, while demand for it grows with every agent anyone deploys. Zatanna has taken one spot in it, the connection layer. There is room for many more, and the companies that take those spots early will look less like AI startups and more like the plumbing of a second internet.

Why it is bigger than it sounds

The short description is integration software. The observation behind it is bigger.

  • The direction of traffic on the web is genuinely changing. Businesses now send agents out through an internet built to invite customers in, and almost nothing exists yet for that direction.
  • Upkeep is the part people keep paying for. Spotting a break and fixing it before the customer notices is exactly the kind of work that renews.
  • Betting that cleverness is not the bottleneck is a clear position. It is also a cheap one to be right about, because every new model release makes it stronger.

What to watch

One number decides this, and it is not requests served. It is what share of connections repair themselves without a human, and how long a break lasts before it is fixed. That single figure is the difference between an infrastructure company and a consultancy with good tools, and it is the only real answer to the long history of things that rot.

The second is concentration. If most of those five billion requests come from a handful of portals and a handful of customers, this is a promising integration business. If the spread is wide, and really does include systems the company had never seen before a customer brought them, then it is closer to the on-ramp for a direction of the web that has barely been built.