Enterprises that want to make and score agents also need to embrace another reality: agents are not made like other software.

Agents are “clearly different” how they are made, how they work, and how they improve, according to it Author CEO and co-founder May Habib. This means digging traditional software development life cycle while dealing with adaptive systems.

“Agents do not firmly follow the rules,” Habib said while staying on stage on Wednesday VB conversion“They are the result-operated. They interpret. They adapt. And behavior actually emerges in the real-world environment.”

Knowing what works-and what does not work-comes from the experience of Babib that helps in the manufacture of hundreds of enterprise clients and to score enterprise-grade agents. According to Habib, more than 350 writers of Fortune 1000 are customers, and more than half of the Fortune 500 will be scaling agents with the author by the end of 2025.

Using non-absent techniques to produce powerful outputs can also be “really nightmares”, Habib said-especially when trying to score agents systematically. Even if enterprise teams can spin agents without product managers and designers, Habib feels that the “PM mentality” is still necessary to cooperate, build, repeat and maintain agents.

“Unfortunately or fortunately, depending on your perspective, it is being quit to hold the bag if they do not take their business counterparts in that new way of the building.”

Why are the target-based agents the right approach

One of the changes in thinking involves understanding the result-based nature of agents. For example, he said that many customers request agents to help their legal teams review or revive contracts. But it is very open. Instead, a target-oriented approach means designing an agent to review and reduce the time of re-formation of contracts.

“Traditional software development life cycle, you are designed for a determined set of very approximate steps,” said Habib. “It is in input, input in a more deterministic manner. But with agents, you are trying to shape the agent behavior. So you are looking for less than a controlled flow and to give a lot to refer to and guide the decision by the agent.”

Another inter -inter agent is creating a blueprint that instructs them with business logic, instead to provide them with workflow. This involves collaborating with subject experts to map designing region loops and promoting processes promoting desired behavior.

While there are a lot of things about scaling agents, the authors are still helping most customers in one construction at a time. This is because it is important to answer questions about the owner and audit of the agent, first of all it is important, which ensures that it remains relevant and still investigates if it is still producing the desired results.

Habib said, “There is a scaling rock that people get very quickly, very quickly without a new approach to the manufacture and scaling of agents.” “There is a rock that people are going to achieve when their organization’s ability to manage agents actually crosses the speed of the development department by the department.”

QA for Agent vs Software

Quality assurance for agents is also different. Instead of an objective checklist, agent assessment involves accounting for non-binary behavior and assess how agents act in real-world conditions. This is because the failure is not always clear – and not as black and white in the form of testing when something breaks down. Instead, Habib said whether it is better to check if an agent has treated well, whether the failure has worked, evaluating the results and intentions: “Here is not the goal is perfection, this behavior is confident, because it has a lot of subjects.”

Habib said that businesses that do not understand the importance of recurrence, they “play a continuous game of tennis that simply wears everywhere until they want to play anymore,” Habib said. It is also important for teams to recover with agents being correct and “is” to launch and run fast and run again and again. “

Despite the challenges, examples of AI agents already help in bringing new revenue to enterprise businesses. For example, Habib referred to a major bank, which collaborated with the author to develop an agent-based system, resulting in new customers onboard several product lines to find a new upsel pipeline of $ 600 million.

New version control for AI agents

Agent maintenance is also different. Traditional software maintenance involves checking the code when something breaks down, but Habib said that AI agents require a new type of version control for everything that can shape behavior. This also requires proper governance and ensures that agents are useful over time instead of provoking unnecessary costs.

Because models do not map clearly for AI agents, Habib stated that maintenance includes the examination signals, model settings, tool skimmers and memory configurations. This also means fully executed in inputs, output, regional steps, tool calls and human interactions.

“You can update [large language model] The LLM Prompt and see the agents behave completely differently, even if anything really changed in the guit history, “said Habib.” Model link shifts, recovery indexes update, tools develop APIs and suddenly the same signal does not behave as expected … It may be felt that we may feel that we are debuting ghosts. ,