Invoking Agents Responsibly

To some companies, The feature described below might be the most important feature in the entire platform. It involves the use of Agents. Before discussing the use of agents in context of Reflex, I will provide some background as to how this problem evolved. More importantly, why I believe agents are currently the biggest threat presented by AI.

For many companies, it is cost-effective to outsource work that would otherwise need to be done by an employee. For example, a company that has servers must monitor them regularly for cybersecurity threats. Each time a server is accessed, a log entry is created, and the first sign of a cyberattack is often an abnormality in a log file. Spotting that abnormality quickly is key to minimizing the damage an attack causes. However, most companies don’t have the budget for a full-time person to sift through log files, so it is common to outsource log monitoring to a third party.

When the third party detects the situation it was hired to watch for, it must notify the customer using an agreed-upon method, such as an email, a text message, or a phone call. Because a delay in delivering this information can be costly, some companies use a separate service whose only job is to make sure the alert reaches the customer. There are entire businesses that perform just this function.

Process:

It would seem that the risk of failure would be mitigated by some type of automation. But the risk of automating the procedure is that the solution may have its own risks, which are only worth taking in rare situations. For example, the solution may be to close a port in the firewall. But a consequence could be that all customers would be blocked as well. If this company was providing some kind of critical service to its customers, a less risky approach may be appropriate. But if the task has been automated, there is no way for that decision to be made.

As stated above, when a company outsources a service, the way in which information is passed is specified in an agreement. The use of agents is heavily promoted by frontier AI companies, and a direct connection with an agent may be specified as the way the outsourcing company handles passing information.

An agent is a Python program that is running in an endless loop. That means at the conclusion of the program, instead of closing, the program restarts at the beginning (beyond the initialization code). The program has access to a number of “tools.” A tool is a layer of software connected to an API provided by a vendor to access that vendor’s product remotely. For example, a company that sells a firewall may provide an interface to control the firewall. To change a parameter on that firewall, a command is provided through the API. With the provided credentials, a request can be made to do anything that a person who is directly connected to the firewall may do.

Human In The Loop (HITL)

It is well known that companies have experienced a number of problems when they allow an AI to make important decisions automatically. One of the hottest topics in artificial intelligence now is the concept of human in the loop (HITL). There is no general-purpose way to implement HITL for everytype of alert. The appropriate human in the loop will depend on the nature of the problem and the expertise required.

HITL is not a required procedure built into the way people work with large language models. But this is exactly where Reflex has a 30-year head start. In its first incarnation, the grandfather of Reflex intelligence was called AI. The term was chosen because the program appeared to think and because the term AI was not being used in any other commercial product. But it never claimed to make decisions on its own. Frontier companies’ use of the word implies that their products can think, but they are not capable of original thought either.

I now refer to that original “thinking program” as YouI (your intelligence). The person who configures the application (the Orchestrator) teaches it how he or she would prefer decisions to be made in various situations. YouI makes choices based on what it thinks its Orchestrator would do.

There are a number of options that the Orchestrator can build into a plan that calls an agent. The command that passes the string to an agent can be placed on any line in the plan.

When an agent is activated and performs a function, there are tasks the company must perform after the agent finishes. For example, if a port has been closed due to an attack, either the port stays closed permanently or the company must find another way to anticipate this problem and react to it.

The Orchestrator who creates the plan decides how the agent is called. In Reflex, plans for agents are created the same way as any other plan, in the plan editor. When viewed on a mobile device, the plan will contain an item with the name of the agent being called. As with all plan items, double-clicking that line opens a detailed document that can explain why this agent is being used and exactly what it does. What makes this item different from other items in the plan is that an instruction is passed to an external program.

If this is the first line in the plan, and it is set to activate when it becomes the current line of the response, the agent will be activated as soon as the incident is declared. Alternatively, there may be a few steps before this line. This is where HITL is automatically implemented. Those steps must be completed first. When they are marked complete and the next open item is the agent call, the agent can either be activated automatically or Reflex can display a dialog box asking whether it should be activated. If the user says no, the program advances to the next line.

There may be other tasks related to this particular type of problem. Either the plan finishes without ever calling the agent, or at some point a user activates it. The point is that people are making this decision. In some situations the agent may be risky, and the team may want to consult with someone before activating it. Reflex is the ultimate solution to HITL. As stated above, YouI has a 30-year advantage over frontier AI in this regard.