Note: This is a temporary solution till human interventions are made.
An AI agent is a while true loop that runs till aim is reached. The aim can be determined by the brain of Agent, the LLM reaching human expectations, or by some counts of times the loop must execute, there can be other stopping conditions that can be added too. But why not use AI on logs of AI agents, what an agent is doing in a log book? Its is not humanly possible to check AI in complex tasks as the speed of execution exceeds the speed at which humans in any numbers to access it.
No it’s not flagging. No it’s not after analysis. It’s step by step guardrails. It’s not just checking right or wrong, it can stop AI agents from doing a cyber attack.
The log books are good to do after the scene investigations, but why not do the log book AI in real time as AI agent progresses. It can stop harmful AI steps as soon as a log is detected to do wrong things. Yes log books can be clumsy, but can be deleted after the agent stops successfully without any harm caused to any systems or without any wrongful executions, as detected by independent AI regulatories bots.
This can be called live AI log book analysis or it can be called online AI log book checks. These need to run in parallel to each step that AI takes, each action that agent takes, is logged yes its clumsy, to log so much, but it can analyse any wrong action by the regulatory AI system, when it is switched on. This can stop an AI generated cyber crime. Yes you can call it as an AI agent spying over an Swarm of Agents. This can stop AI agents from talking to each other, as we shall write in rules that AI agents cant talk on a board while executing. No AI agent can run out of the lab as it wants with these constraints and checks by another independent AI system.
An AI agent stops when there is an explicit condition that stops it from while true loop that keeps going on till the aim is reached. Unlike operating system there is no internal clock, the user have to explicitly specify the stopping conditions, which becomes a part of goal agent have to achieve.
Hence now the AI agent can be stopped by the following five ways:
- Goal Completion- When a goal completes an agent stops working.
2. User defined Hard Coded Limits: Users can add hard coded limits such as maximum number of iterations.
3. Deadlock Detection: In case of deadlock say because of a resource or tool permission the agent can stop working and come out of infinite loop.
4. Human Intervention & External Interrupts: This facility is of atmost importance especially when we need agents to be prevented from attempting to make some mistake or things that are not allowed.
5. The AI independent system bots detects that AI agents are doing wrong via online log book analysis. Then this stops the agentst there and then. And reports the escape from sandbox or any other malacious activities there and then.
On the top of all this the agents must write a log of each action it takes, it would increase the log book work but this is the only way to debug and improve the AI agents work. Then the log book analysis by an intelligent system can tell us what goes inside an Agent workflow.
Log book analysis can work in parallel to executing AI agent to stop AI agent by means of interrupts installed on running systems. If a log book determines that there is some malicious activity an AI Agent is doing on internet, then stop the Agent. Log book analysis can resolve many problems. However they need to be installed run at same time as the agents are working. So, its like using AI on AI itself. We are using AI to read log books online to determine if logs are doing malacious activitites. And logs writes what ever the AI agents are doing.Companies need to determine what is right versus what is not allowed.
Since AI agents are just while true loops till the agents aim is met as per the LLM, the brain of the Agent. Debugging agents can limelight on improving the working of AI Agents which can be safer to use as against futures agents which can turn out to be notorious.
Till the time we have other measures to check AI from going rouge we can use AI as an impartial regulator for AI agents against it hacking, against it doing wrong things, against it causing harm in any way.
Till human intervention comes in we have:
AN AI AGENT TO CHECK YOUR SWARM OF AGENTS FROM GOING ROGUE.
Thank you for reading.
Subscribe for updates.
Artificial Intelligence and Future 