Let’s start with the term “agent” yourself. Right now, it is being slapped on everything from easy script to sophisticated AI workflows. There is no joint definition, which makes the main automation marketing for companies a lot of space, because it is far more developed. Such “agent washing” does not just confuse consumers. It invites frustration. We do not necessarily need a strict quality, but we need clear expectations about these systems to do what these systems do, how much they work, and to what extent they perform reliable.
And the next major challenge is vigorous. Most of today’s agents are powered by a large language model (LLM), which produces potential reactions. These systems are powerful, but they are unexpected. They can prepare things, get off track, or fail in subtle ways – especially when they are asked to complete the multi -stipped task, drag outer tools and make the LLM response a chain together. A recent example: An automated support agent of a famous AI programming assistant, cursor, told an automatic support agent that he could not use software on multiple device. There are widespread complaints and notifications to cancel consumers. But it turned out The policy did not exist. The AI ​​invented it.
In enterprise settings, this type of error can cause a lot of damage. We need to prevent LLM from being treated as stand loan products and need to start construction of a complete system around them – the systems that cause uncertainty, monitor outputs, monitor costs, manage costs and layers for safety and accuracy. These steps can help ensure that the output is following the requirements expressed by the user, adheres to the company’s policies on access to information, respects privacy issues, and respects. Some companies, including AI11 (which I sponsored and who has received funds from Google), are already going in this direction, and more deliberately, the language model in structural architecture. Our latest launch, a teacher, is designed for enterprise reliability, which can ensure reliable output in combination with company data, public information and other tools.
Nevertheless, even the smartest agent will not be useful in space. Agent Model Working Different, various agents need to cooperate without permanent human surveillance (your travel booking, weather testing, reporting your expenses). This is where Google’s A2A protocol comes. Its purpose is a universal language that allows agents to share what they can do and distribute tasks. In principle, this is a good idea.
In practice, A2A still comes short. This explains how the agents talk to each other, but not it means. If an agent says he can provide “air conditions”, the other has to guess whether it is useful to test the weather on the flight route. Without shared words or contexts, coordination easily breaks. We have seen this issue in the first divided computing. Solving it on a scale is far from modest.