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

AI agents have security and reliability issues. While agents will allow businesses to automate more steps in their workflow, they can take unexpected actions when performing tasks, not very flexible and difficult to control.

Organizations have raised alerts about unreliable agents, fearing that once deployed, agents may forget to follow instructions.

Openai Even admitting that ensuring agent reliability will involve working with external developers, so Opened its agent SDK Help solve this problem.

However, researchers at the Singapore Management University (SMU) have developed A new method Solve proxy reliability.

AgestentsPec is a domain-specific framework that allows users to “define structured rules that combine triggers, predicates and law enforcement mechanisms”. The researchers say agent PEC will only work in parameters that users want.

Guiding LLM-based agents through new methods

AgentsPec is not a new large language model (LLM), but a method to guide LLM-based AI agents. Researchers believe that agents can be used in enterprise environments and in autonomous driving applications.

Integrated in Langchain The framework, but the researchers say they designed it as a framework – agile, which means it can also run on the Autogen and Apollo ecosystems.

Experiments using proxy PEC show that it prevents “more than 90% of unsafe code execution, ensures full compliance with autonomous driving legal aggression scenarios, eliminates dangerous actions in the reflected proxy tasks, and develops with millisecond-level elevated development.” O1, LLM generated agent PECENTSPEC rules also have strong performance and executes 87% of the risk code and prevents “breach of the law in 5 of 8 cases.”

The current method is a bit lacking

Proxy PEC is not the only way to help developers provide more control and reliability for proxy. Other methods include tools and guards. Start a business Galileo emission Agent evaluationa way to ensure agents work as expected.

Open source platform H2O.AI uses predictive models Improve the accuracy of the agents used by companies in finance, healthcare, telecommunications and government.

Current risk mitigation methods, such as tool emu, can effectively identify risks, said agent Pec. They noted that “these methods lack interpretability and do not provide mechanisms for safe execution, making them vulnerable to adversarial manipulation.”

Using a proxy

AgestentsPec is the runtime execution layer of the agent. It intercepts the agent’s behavior while performing a task and adds security rules that are set or prompted by humans.

Since AdventsPec is a language for custom domains, users must define security rules. There are three components to this: the first is the trigger, when it activates the rule; the second is to check the added conditions. The third is to execute, if the rules violate, the action is performed.

But, as previously mentioned, AgentsPec builds on Langchain, and researchers say agents can also integrate into other frameworks, such as Autogen or Automons weather Software Spack Apollo.

These frameworks coordinate the steps the agent needs to take, by taking user input, creating an execution plan, observing the results, and then determining if the operation is completed and if not, plan the next step. AgentsPec adds rule execution to this process.

Before performing an action, AdentsPEC evaluates predefined constraints to ensure compliance and modify the agent’s behavior if necessary. Specifically, an agent hooks Adestspec in three key decisions: after performing an action (agent action), after performing (agent), after performing an observation (agent), and when the agent completes its tasks (agent), without the core of these points.

More reliable agents

Approaches like AgesentPec emphasize the need for reliable agents used by enterprises. As the organization begins Plan its proxy policythe technical decision-making leader also considers ways to ensure reliability.

For many, agents will eventually automatically and proactively perform tasks for users. this The idea of ​​environmental agent,If AI agents and applications are constantly running in the background and triggering their own actions, they need agents that do not displace from their paths and accidentally introduce non-safe actions.

If the environmental agent is a proxy AI that will be done in the future, methods like AdmentPec will spread as companies seek to make AI agents consistently reliable.