Hidden scaling cliff that is going to break the implementation of the agent

Enterprises that want to build and scale agents must also adopt a different reality: agents are not built like other software.

Agents are “categorically different” in terms of construction, how to work and improve them, according to this Writer CEO and co -founder can habib. This means abandoning the established life cycle of software development in the case of adaptation systems.

“Agents do not reliably follow the rules,” said Habib on Wednesday on the stage at VB Transform. “They are based on the results. They interpret. They adapt. And behavior really only appears in real environments.”

Knowledge about what works-what does not work-he loves the experience of habib, helping hundreds of corporate clients in building and scaling the scale of corporate class agents. According to Habib, over 350 of the 1000 fortune are writers, and more than half of the 500 fortune will be scaling agents with the writer by the end of 2025.

Habib said Habib, the apply of innocent technology to produce powerful outputs can be “really nightmarish”-especially during systemic scaling of agents. Even if corporate teams can return agents without product managers and designers, Habib believes that “PM thinking” is still needed for cooperation, building, iteration and maintenance of agents.

“Unfortunately, or fortunately, depending on your perspective, it will remain, holding the bag if they do not lead their business counterparts in this new way of building.”

Why agents oppose goals are the right approach

One of the changes in thinking includes understanding based on the result of the nature of agents. For example, she said that many clients ask agents for aid to their legal teams in browsing or redlining contracts. But it’s too open. Instead, a goal -oriented approach means designing an agent to shorten the time spent on review and reduce contracts.

“In the traditional life cycle of the software you design for a deterministic set of very predictable steps,” said Habib. “There is a contribution, a contribution to a more deterministic way. But together with agents you try to shape agency behavior. So you are looking for less controlled flow and much more to give the context and conduct a decision by an agent.”

Another difference is to build a plan for agents that instruct them with business logic instead of providing them with flow flows to follow. This includes the design of the reasoning loop and cooperation with entities experts in order to map the processes promoting the desired behavior.

While there is a lot of talk about scaling agents, the writer still helps most customers in building them individually. This is because it is first significant to answer questions about who owns and controls an agent who makes sure that he remains appropriate and continues to check whether he still gives the desired results.

“There is a scaling cliff, to which people reach very, very quickly without a new approach to building and scaling agents,” said Habib. “There is a cliff that they will come to when the ability to organize them to manage agents responsibly exceeds the pace of development of the department acting by the department.”

QA for agents versus software

Quality assurance is also different for agents. Instead of an objective control list, the agency assessment includes the settlement of non-bining behavior and assessing how agents operate in real situations. This is because the failure is not always obvious – and not as black and white as checking if something broke. Instead, Habib said that it is better to check if the agent behaved well, asking if the emergency rules worked, assessing the results and intentions: “The goal here is not perfection, is behavioral confidence, because there is a lot of subjectivity in this.”

Companies that do not understand the importance of iteration end in “continuous tennis game, which simply consumes every side until they want to play,” said Habib. It is also significant that the teams are fine when the agents are less than perfect, and more about “safe starting their quick start and iteration.”

Despite the challenges, there are examples of AI agents that already aid bring fresh revenues for enterprises. For example, Habib mentioned a immense bank that cooperated with the writer to develop a system based on agents, which results in a fresh UPSell pipeline worth $ 600 million by implementing fresh customers into many product lines.

The fresh version controls AI agents

Agency maintenance is also different. Time-honored software maintenance is about checking the code when something breaks, but Habib said that AI agents require a fresh type of control of the version of everything that can shape behavior. It also requires adequate management and assurance that agents remain useful over time, instead of incurring unnecessary costs.

Because the models do not map cleanly to AI agents, Habib said that maintenance includes checking of prompts, model settings, tool diagrams and memory configuration. It also means full tracking between entrances, outputs, reasoning, tools and interpersonal interactions.

“You can update [large language model] LLM prompt and watch the agent behaves completely differently, although nothing in history GIT has actually changed, “said habib.” The model combines changes, the download indexes will be updated, interpret the API of tools Ap.