Many organizations will feel reluctant to restore their tech stack and start from the beginning. Not Idea. For the 3.0 version of its production software (released in September), the company did not hesitate to rebuild the ground -up. He recognized that in fact, it is necessary to support Agent AI on the enterprise scale. Although traditional AI -powered workflows include clear, phased guidelines that are based on some shot learning, the AI agent running through advanced reasoning models is thinking about the definition of the device, which tools they have and can plan the next steps. “Instead of trying to re -work in what we are making in what we are making, Sarah Sax Bat told Venture Bat, head of the AI modeling concept, we wanted to play the powers of the models of reasoning.” “We have rebuilt a new architecture because work is different from the floose agents.”
Re -or -or -or -to -model so that the models can work independently
The idea is adopted by 94 % of the Forbes AI50 companies, it has 100 million total consumers and its users are counted in the curses, cursors, Figma, Ramp and Versel. Rapidly developed AI landscapes, the company identified the need to move towards a round -based reasoning system beyond the task flu, which allow agents to select autonomy in the linked environment, implement the arcket and tools.
Sex said very quickly, learning to use models tools of reasoning and following the China -off thinking (COT) instructions have become “better”. This can help them become “far more free” and make several decisions within an agent workflow. “We rebuilt our AI system to play with it," He said. Sex explained that from the point of view of engineering, this means to replace the strict instant flow with the united orchestration model. This basic model is supported by modular sub -agents who find theory and add to the web, inquiry and database and edit content. Each agent uses the tools with context. For example, they can decide whether to find ideas themselves, or another platform like a silic. Unless the model is found relevant information, the search will be conducted after another. Then, for example, the notes can turn into tips, follow -up messages, track task, and spots, and update to knowledge bases. In concept 2.0, the team focused on performing a specific task to the AI, under which the models needed to “think fully” about ways to indicate. However, with version 3.0, user can assign work to agents, and agents can actually take action and perform multiple tasks simultaneously. “We have re -chosen to choose itself on the tools to work again, which is clearly a way to go through all these different scenarios,” said Sex explained. The purpose is to ensure the interface of everything with AI and that “whatever you can do, you can think agent.”
To divide to be isolated
The philosophy of “better, faster, cheap” concepts a permanent repetition that balances delays and accuracy through fine tonic vector embedded and flexible search correction. The sex team uses a rigorous diagnostic framework that combines the test, language correction, human infrastructure data and LLMS-A-PID, which identifies contradictions and errors in model-based scoring. Sex explained, “By dealing with the diagnosis, we are able to identify where the problems come from, and it helps us to isolate unnecessary deception.” In addition, making architecture itself means that it is easy to make changes with the evolution of model and technique. Sex noted, “We improve the maximum delay and parallel thinking,” which leads to “better accuracy”. The model is set up in the web and concept -connected workplace data. Finally, Sex reported, the investment in its architecture rebuilding has already provided conceptual profits in terms of rapid rate of capacity and change. “When we make the next progress, we are fully open to rebuild it, if we have to do,” he added.
Understand the delay in context
When making models and creating fine toning models, it is important to understand that the delay is the source: AI must provide the most relevant information, not necessarily, at the speed cost. “You will be surprised at different ways that consumers are willing to wait for things and wait for things,” said Sex. It has an interesting experience: How slow can you be before leaving the model? With pure navigational search, for example, consumers cannot be so patient. They want nearly answers. Sex pointed out, “If you ask, ‘What are two plus two’, then you don’t want to wait for your agent to find everywhere in Slack and Jira.” But the more time it is, the more reasonable agents can be. For example, the concept can perform 20 minutes of independent work in hundreds of websites, files and other content. In these examples, consumers are more willing to wait, Sex explained. They allow the model to be hanged in the background while participating in other tasks. “This is a product question,” said Sex. “How do we set the user’s expectations from the UI? How do we address the user’s expectations late?”
Idea is its largest user
The idea understands the importance of using their products – in fact, its employees are among its largest electric users. Sex explained that teams have active sandboxes that produce training and diagnosis data, as well as a “really active” thumb down the user’s feedback. Consumers are not ashamed to say whether they think it should be better or the features they want to see. Sex emphasized that when a user thumbs up a conversation, they are clearly allowing a human interpretation to analyze that the conversation does not name this way. “We’re using our tool all day, every day, as a company, and so we get a really fast feedback,” said Sex. “We are really digging our own products.” He said, “This is his own product that he is building, Sex noted, so he understands that when it comes to quality and functionality, they can find fountains.” To balance it, the concept is trusting "Very AI-Savvy" Design partners who are given quick access to new abilities and provide important feedback. Sex emphasized that it is as important as internal prototying. “We’re all about to do open -minded experience, I think you will get a lot of comments,” said Sex. “Because at the end of the day, if we just see how the concept uses the concept, we are not really giving our users the best experience.” As the important thing is, continuous internal testing teams allow to evaluate and ensure that models are not being registered (when accuracy and performance decreases over time). "What you are doing is loyal," Sex explained. "You know that your delay is within the limits."
Many companies make a mistake to focus too much on retroxic concentrated Evans. Sex indicated that it is difficult for them to understand how and where they are getting better. The idea considers Evils as a "Latemus test" Performing development and waiting for development and observation and reactionary proof. “I think many companies are in conflict with them,” said Sex. “We use them for both purposes. We really think about them.”
Tech from the trip of concept
For businesses, the concept can act as a blueprint on how an associated, permit can be implemented responsibly and dynamically agent AI in the enterprise workplace. The way to Sachs for other tech leaders:
When basic abilities change, do not be afraid to rebuild. The idea has completely re -engineered its architecture to align with reasoning models.
Treat delay as a context: Improve the issue of use rather than globally.
All results to ensure accuracy and confidence in reliable, curating enterprise data. He suggested: “Be willing to make tough decisions. Be willing to sit on the upper part of the Frontier, so to talk, what are you developing to make the best product for your customers.”