New approach to agent reliability, AgentSpec, forces agents to follow rules

Artificial intelligence factors have safety and reliability problems. Although agents will allow institutions to automate more steps in the progress of their work, they can take unintended measures while carrying out a task, not very flexible and difficult to control.

Organizations have already sparked alert about unreliable agents, concern that once it is published, agents may forget to follow the instructions.

Openai admitted that the agent’s reliability guarantee will include work with external developers, so its agents opened SDK to help solve this problem.

However, Singapore University researchers (SMU) has developed a new approach to a reliability of the agent.

Agentspec is a special framework that allows users to “define structured rules that include operators, predictions and enforcement mechanisms.” The researchers said that Agentspec will make agents only among the parameters that users want.

LLM agents are directed with a new approach

Agentspec is not a new LLM model, but rather an approach to the guidance of LLM -based artificial intelligence agents. Researchers believe that Agentspec can be used for agents in self -driving institutions and applications.

The first agents tests on Langchain frameworks, but researchers said they designed them to be a frame of a frame, which means that automatic ecosystems and pyroo can also be operated.

The experiments used by CNCENTSPEC have shown that they prevented “more than 90 % of unsafe executions, ensuring complete compliance with independent legal driving scenarios, and removes dangerous procedures in the tasks of the sensor agent and works with costs at the level of millimeters.” Openai’s O1, created by LLM, which used Openai’s O1, had a strong performance and 87 % of risk -framed law and prevent “breaking law in 5 out of 8 scenarios.”

The current methods lack a little

Agentspec is not the only way to help developers giving agents more control and reliability. Other methods include toolemu and Guardagent. Galileo’s startup launched the agent’s assessments, a way to ensure the work of agents in an intended manner.

The H2o.ai open platform uses predictive models to improve the accuracy of agents that companies use in financing, health care, communications and government.

Both researchers said that the current methods of risk alleviation, such as Toolemu, effectively define risks. They pointed out that “these methods lack interpretation and do not provide any mechanism to enforce safety, which makes them vulnerable to numerical manipulation.”

Using agents

Agentspec acts as an enforcement time for agents. He objects to the agent’s behavior during the implementation of the tasks and adds the safety rules that humans have developed or created by claims.

Since Agentspec is a specially dedicated language, users must determine safety rules. There are three components for this: The first is the trigger, which sets the time to activate the rule; The second is to verify to add conditions; The third is the enforcement, which imposes the actions that must be taken if the rule is violated.

Although CNCENTSPEC is based on Langchain, previously mentioned, the researchers said that Agentspec can also be combined into other frameworks like Autogen or SostMon Software Stack Apollo.

These frameworks regulate the steps that agents must take by taking the user inputs, creating an implementation plan, and monitoring the result, then determining whether the procedure has been completed, and if not, planning the next step. Agentspec adds the enforcement of the base to this flow.

“Before carrying out a procedure, Claspik evaluates the pre -determined restrictions to ensure compliance, and adjusts the agent’s behavior when necessary. Specifically, each of the Cancentspec connects it to three main decision points: before implementing the procedure (Agentance), these points stipulate the structure”, unchanged, and change them.

More reliable factors

Methods such as Agentspec emphasize the need for reliable factors for the organization’s use. As institutions start planning their agent strategy, technical decision leaders are also looking into ways to ensure reliability.

For many, agents will eventually carry out independent and independent tasks for users. The idea of ​​the surrounding agents, where artificial intelligence agents and applications work continuously in the background and operate themselves to implement the procedures, from agents who do not scream from their path and mistakenly provide unsafe procedures.

If the surrounding factors are where Agentic AI will go in the future, I expect more methods like Ogentspec to multiply with companies to make artificial intelligence agents constantly reliable.