Agents that actually work

Why this conversation

An agent that works inside a company’s systems raises a question every team meets sooner or later: how do you know the agent is acting for the right person, and that it cannot do more than that person could? Nathan Slavik leads content at Arcade.dev. The host recorded with him at AGNTCon + MCPCon North America in San Jose on Thursday 22 October 2026, and Nathan answered with three steps, a story about booking time off, and the host’s own inbox as the example. The rest of that day is in AGNTCon + MCPCon North America 2026: day one.

Authorization, enforcement, governance

Asked what Arcade.dev does, Nathan said that it “makes agents actually work”, and he split that into three steps.

The first is authorization: is this agent really allowed to act on behalf of this user? His example was someone in a business who wants an agent to sign into Salesforce. The question then is how anyone can be sure the agent is signing in as that person.

The second is enforcement. Once inside Salesforce, the agent should be able to do exactly what its user can do, with no extra rights and none missing. If the user has not allowed it to delete a customer record, it should not be able to.

The third is governance, meaning a record of everything the agent does. With a full audit trail, anyone looking back can tell which agent acted, for which person, what it did and when. Nathan says it is the combination of those three layers that leads large banks to trust Arcade inside their systems.

From engineering teams to HR

The biggest use case, as at so many businesses in the industry, begins with engineering and dev teams. Once Arcade is in their systems, it spreads from there. Much of the traction Nathan sees today is with GTM teams, and the company learned quickly about a use that may sound simple: HR. Most employees would rather not log into Workday at all. They would rather tell their agent they will be off for a week and let it handle the scheduling. The same questions apply: the agent has to reach Workday securely and book the time on the employee’s behalf.

An action runtime in the middle

The host asked whether every employee gets a digital companion, or whether there is one general agent that everyone shares. Nathan’s answer was that Arcade takes the permissions and policies a company already has and translates them to its agents. To make it concrete, he asked the host which program he uses all the time; after Slack, they settled on Gmail. Arcade then sits between the host and Gmail as an action runtime. When the host tells his agent to send an email, Arcade first checks that this really is his agent, and the agent goes into Gmail. When the moment comes to send, the request passes back through Arcade, which checks that the agent may send it. Nothing happens, at any time, without the user’s approval.

Any model, any number of agents

What if a new model comes out that the host wants to try? Arcade is completely model agnostic, Nathan says: whether a team works with Claude, Cursor, Grok or something else, it can connect what it uses to Arcade.

Nor is there a limit on the number of agents. In his view, “the more that you use Arcade, the more benefit you see”. A small team with five employees and ten agents might decide it can keep an eye on them by hand. A business with a hundred or a thousand employees, each with ten agents connecting to 50 business systems, is another matter: “That’s when you need Arcade to be able to automatically check all of those policies”. Otherwise someone would have to sit and check by hand every time anyone wanted to use an agent, and that would take ages.

To try it in an organisation, go to arcade.dev and sign up for free, which gives access to the basic tools and to the best of the platform. Enterprises that want to run Arcade on premises or air-gapped, or that have other special needs, can reach out to the team to be set up. The host closed by proposing to build the AI agent of tomorrow; Nathan suggested doing it today.

When it comes out

The episode is edited and in the release order. Like every episode it goes out on YouTube first, then on Apple Podcasts, Spotify and Amazon Music.