New approach for the reliability of the agents, agent peaks, forced agents, to comply with rules

AI agents have security and reliability problems. Although agents would allow companies to automate more steps in their work processes, they can take unintended actions while performing a task, are not very flexible and difficult to control.

Organizations have already triggered the alarm via unreliable agents and fear that agents may forget to follow instructions after their use.

Openai Even admitted that the guarantee of the reliability of the agents would include cooperation with external developers. opened his agents SDK To solve this problem.

However, the researchers at Singapore Management University (SMU) have developed A new approach Reliability of the agents.

Agentspec is a domain -specific framework that can “define structured rules that contain trigger, predicates and assertiveness”. The researchers said that Agentspec would make agents only work within the desired parameters.

Management of LLM -based agents with a new approach

Agentspec is not a new LLM, but an approach to lead LLM-based AI agents. The researchers believe that AgentsSpec can not only be used for agents in corporate environments, but also useful for self -driving applications.

The first Agentspec tests that are integrated Praise Frameworks, but the researchers said they had designed it as a framework-agnostic, which means that it can also be carried out on ecosystems on autogenic and Apollo.

Experiments using Agentspec showed that “over 90% of the execution of uncertain code designs was prevented, the scenarios for autonomous driving ensures the laws to combat illegality, remedy dangerous actions in embodied agent tasks and operated with overhead to millisecond levels”. The LLM-generated Agentspec rules that opened O1 also used, also achieved a strong performance and forced 87% of the risky code and prevented “laws in 5 of 8 scenarios”.

Current methods are missing a little

Agentspec is not the only way to help developers bring agents more control and reliability. Some of these approaches include Toolemu and Guardagent. The startup Galileo launched Agent ratingsA way to ensure that agents work as intended.

The open source platform H2o.ai uses predictive models To make agents of companies in the areas of finance, healthcare, telecommunications and government more precise.

The AgentsSpec said that researchers said current approaches to reduce risks such as Toolemu effectively identify the risks. They found that “these methods have no interpretability and do not offer a mechanism for enforcing security, which makes them vulnerable to controversy manipulation.”

Use of Agentspec

Agentspec works as a runtime -enforcement layer for agents. The agent’s behavior begins to carry out tasks and add security rules that are defined by humans or generate by input requests.

Since Agentspec is a custom domain -specific language, users must define the safety rules. This gives three components: the first is the trigger that represents when the rule is to be activated; The second is to check to add and enforce conditions to take the measures when the rule is violated.

Agentspec is based on Langchain. However, as already mentioned, the researchers said that Agentspec can also be integrated into other framework conditions such as autogenic or autonomous vehicle software -stack Apollo.

These frameworks orchestrating the stages that agents have to take by taking up the user input, creating a execution plan, observing the result, and then decides whether the action has been completed and if not, this plans the next step. Agentspec adds this river to the rule.

“Before an action is carried out, Agentspec evaluates predefined restrictions to ensure compliance with compliance and to change the behavior of the agent if necessary. In particular, the Agentspec hooks in three key decisions: Before an action is carried out (agentation), an action creates an observation (agentstep), and if the agent completes the task (agent finish).

More reliable agents

Approaches such as Agentspec underline the need for reliable agents for the use of companies. Start as organizations Plan your agent strategyTech decision leaders also check for ways to ensure reliability.

For many, agents will ultimately perform tasks for users autonomously and proactively. The Presentation of ambient agentsWhere AI agents and apps continuously run in the background and execute to carry out actions, agents must require that do not differ from their way and not to introduce secure actions.

If environmental agents will go in the future of agents -KI, they expect more methods such as Agentspec, since companies want to continuously make the AI ​​agents reliably reliably.