Why do not most AI agents reach the institutions to production and how do Databricks plan to fix it

The development efforts of the AI ​​Enterprise Ai agent never reach production and not because technology is not ready. The problem, according to DatabricksIt is that companies still rely on manual assessments through a slow, inconsistent and difficult process.

Today in Data + AI Summit, Databrics launched bricks the mosaic agent as a solution to this challenge. Technology depends on the expansion of the domain Ai mosaic agent framework framework The company announced in 2024 simply, it is no longer good enough to be able to build artificial intelligence agents in order to have a real impact.

The Mosaic Agent Bricks platform to automate the agent using a series of research -backed innovations. Among the main innovations is merging Tao (Adaptive improvement time), which provides a new approach to controlling artificial intelligence without the need for the name called. The mosaic agent also creates artificial data for the field, creates the task -realized criteria and works to improve the balance between quality to the cost without manual intervention.

Basically, the goal of the new statute is to solve a problem facing Databrics users with efforts to develop the current artificial intelligence agent.

“They were blinded, and they had no way to evaluate these agents,” Hanlin Tang, chief technology official in nerve networks, told Venturebeat. “Most of them depended on a type of manual manual tracking to see if the agent looks good enough, but this does not give them confidence to enter production.”

From research innovation to AI Enterprise production scale

Tang was previously co -founder and CTO of the mosaic, which was It was acquired by Databricks In 2023 for $ 1.3 billion.

In Mosaic, innovation in the field of research was not necessarily an immediate institutional impact. All this changed after the acquisition.

“It was the moment of the large lighting bulb for me when we first launched our products on Databrics, and immediately, overnight, we had, like thousands of the agents of the Foundation who used it,” Tang said.

In contrast, before the acquisition, mosaics spend months in an attempt to get a few institutions to try products. The mosaic integration in Databrics gave the mosaic research team direct access to the problems of institutions on a large scale and revealed new areas for exploring them.

The Foundation’s contact has revealed new research opportunities.

“It is only when you have a connection with Enterprise customers, you are working with them deeply, you are already revealing a kind of interesting research problems to pursue them.” “Brick client …, in some aspects, a kind of evolution of everything we were working on in Mosaic now after we were all fully brick.”

Solve the crisis of evaluation of artificial intelligence factors

Foundation teams face an expensive and accurate improvement process. Without standards of knowledge of the task or the test data of the field, each modification of an expensive guessing game agent becomes expensive. Quality of drifting, exceeding the costs and permanent deadlines for follow -up.

Agent Bricks automate the entire improvement pipeline. The statute takes a high -level task description and institution’s data. The rest is treated automatically.

First, it generates assessments of the mission and LLM judges. Next, it creates artificial data that reflects customer data. Finally, looking through improvement techniques to find the best composition.

“The customer describes the problem at a high level and does not enter the details of the low level, because we take care of it,” Tang said. “The system creates artificial data and builds LLM judges for each task.”

The statute provides four agent compositions:

• Information extractionDocuments (PDF, emails) transform into organized data. One of the cases of use can be the retail institution that you use to withdraw the details of the product from the suppliers PDFS, even with complex format.
• Knowledge assistantProvides accurate answers and cited from the institution’s data. For example, manufacturing technician can obtain immediate answers from maintenance guides without drilling via folders.
• Llm dedicatedText transforming tasks (summary, classification). For example, health care institutions can customize models that summarize the patient’s notes for clinical work.
• Multi -agent supervisorIt regulates multiple factors for the complex workflow. One example of use is financial services companies that can coordinate agents to discover intention, restore documents and check compliance.

Agents are great, but do not forget the data

Building and evaluating agents is an essential part of AI Enterprise, but it is not the only part required.

Databricks is placed from mosaic brick as an artificial intelligence consumption layer sitting over a unified data staple. In Data + Ai Summit, Databricks also announced its general availability Lakeflow Data Engineering A platform, which was first examined in 2024.

Lakeflow solves the data preparation challenge. It unites three important data engineering trips that had previously required separate tools. Swallowing in obtaining organized and non -organized data in data data is compatible. The shift provides effective, reshape and prepared data. It manages the progress of production and schedule.

Workflow connection: Lakeflow is the institution’s data through swallowing and unified transformation, then Agent Bricks builds improved artificial intelligence agents on these prepared data.

“We help enter data into the statute, and then you can do ML, BI and AI analyzes,” Bilal Aslam, Director of Products Management at Databrics told Venturebeat.

Except data swallow, the mosaic agent brick also benefits from Databrics’s governance features. This includes arrival control and data tracking elements. This integration guarantees that the agent’s behavior respects the governance of the institution’s data without additional formation.

The customer learns from human comments eliminates immediate filling

One of the common methods for directing artificial intelligence agents today is the use of system router. Tang referred to the exercise of a “wavy filler” where users provide all kinds of guidance to a mentor in the hope that the agent will follow.

Agent Bricks introduces a new concept called – learning the agent from human reactions. This feature automatically adjusts the system components based on natural language instructions. It solves what Tang calls the instant stuffing problem. According to Tang, the immediate filling approach often fails because the agent’s systems have multiple ingredients that need to be modified.

Learning the agent from human reactions is a system that automatically explains the natural language instructions and control the appropriate system components. The approach reflects learning to enhance human reactions (Rlhf) But it works at the level of the agent instead of the individual model weights.

The system takes two main challenges. First, natural language instructions can be mysterious. For example, what does “respect for your brand sound” really mean? Second, the agent systems contain many training points. The difference is struggled to determine the ingredients that need to be modified.

The system cancels guessing around the components of the agent that needs to be modified for specific behavioral changes.

“This will believe that it will help the agents to become more severe,” said Tang.

Technical advantages on the current frameworks

There is no shortage of work frameworks and agent of artificial intelligence on the market today. Among the growing list of sellers options tools from Linjshenand Microsoft and Google.

Tang argued that what makes the mosaic agent bricks different is improvement. Instead of requesting and adjusting manual composition, working bricks merge multiple search techniques automatically: Tao, learning within context, immediate improvement and control.

When the agent comes to the agent’s contacts, there are some options on the market today, including Google’s 2age agent protocol. According to Tang, Databrics is currently exploring the various agent protocols and has not committed to one standards.

Currently, the brick customer deals with the agent to the agent through two basic ways:

1. Exposing agents as end points can be wrapped in different protocols.
2. Using a multi -agent supervisor who realizes MCP (the context of the context of the model).

The strategic effects of decision makers of institutions

For institutions looking to lead the road in artificial intelligence, it is very important that the correct techniques are in place to evaluate quality and effectiveness.

The publication of agents without evaluation will not lead to an ideal result and will not have agents without the basis of solid data. When considering agents development techniques, it is important to have suitable mechanisms to assess the best options.

It is also noted that the agent he learns from the human feedback approach is worth noting for the institution’s decision makers because it helps to direct Aic AI to the best results.

For institutions looking to lead in the deployment of an artificial intelligence agent, this development means that the evaluation infrastructure is no longer a ban. Institutions can focus resources on the use of a condition to determine the condition and prepare data instead of building improvement frameworks.