The hidden scaling cliff that’s about to break your agent rollouts

Companies wishing to build and evolve agents must also adopt another reality: agents are not built like other software.

The agents are “categorically different” in the way they are built, how they work and how they are improved, according to Writer CEO and co-founder May Habib. This means abandoning the traditional software development life cycle when it comes to adaptive systems.

“Agents do not respect reliably,” said Habib on stage on stage on Wednesday VB transform. “They are focused on results. They interpret. They adapt. And behavior really only emerges in real environments. ”

Know what works – and what does not work – comes from Habib’s experience to help hundreds of business customers to build and set up business quality agents. According to Habib, more than 350 of fortune 1000 are customers of writers, and more than half of fortune 500 will make the scale of agents with writer by the end of 2025.

The use of non -deterministic technology to produce powerful outings can even be “really nightmarish,” said Habib – especially when you try to scale up agents systematically. Even if business teams can run agents without product managers and designers, Habib thinks that a “PM state of mind” is always necessary to collaborate, build, iterate and maintain agents.

“Unfortunately or fortunately, depending on your point of view, they will stay holding the bag if they do not drive their corporate counterparts in this new way of building.”

>>See all of our Transform 2025 coverage here<

Why are the objectives -based agents the right approach

One of the reflection changes includes understanding of nature based on the results of agents. For example, she said that many customers were asking agents to help their legal teams examine or redlong contracts. But it’s too open. Instead, an approach focused on objectives means designing an agent to reduce the time spent examining and redlinging contracts.

“In the life cycle of traditional software development, you design for a deterministic set of very predictable steps,” said Habib. “It is an entry, an entry in a more deterministic way. But with the agents, you are looking to shape agenic behavior. You therefore look for less a controlled flow and much more to give the context and guide decision -making by the agent. ”

Another difference is to build a plan for agents who educate them with commercial logic, rather than providing them with workflows to follow. This includes the design of reasoning loops and collaboration with experts in matters to map processes that promote desired behavior.

Although there are a lot of discussions on scaling agents, the writer still helps most customers build them at a time. Indeed, it is important first of all to answer the questions on who owns and audits the agent, who ensures that he remains relevant and always checks if he always produces the desired results.

“There is a cliff on the scale that people arrive very, very quickly without a new approach to construction and scale agents,” said Habib. “There is a cliff to which people will arrive when the ability of their organization to manage agents in a responsible manner really exceeds the rhythm of the department of development of the department by department.”

QA for software VS agents

Quality insurance is also different for agents. Instead of an objective control list, agent evaluation includes taking into account non -binary behavior and the evaluation of the way in which agents act in real situations. It is because failure is not always easy – and not as black and white as to check if something has broken. Instead, Habib said that it was better to check if an agent behaved well, asking if the failures worked, evaluating the results and intention: “The goal here is not perfection, it is behavioral confidence, because there is a lot of subjectivity in there.”

Companies that do not understand the importance of iteration end up playing “a constant tennis game that wears out on each side until they no longer want to play,” said Habib. It is also important that the teams agree with the agents who are far from perfect and “launch them safely and run quickly and iterate again and again”.

Despite the challenges, there are examples of AI agents that are already helping to generate new income for businesses. For example, Habib mentioned a large bank that collaborated with Writer to develop a system based on agents, which has led to a new sales pipeline with a value of $ 600 million by integrating new customers into several product ranges.

New version commands for AI agents

Agent maintenance is also different. Maintenance of traditional software consists in checking the code when something breaks, but Habib said that AI agents require a new type of version control for everything that can shape behavior. It also requires good governance and ensuring that agents remain useful over time, rather than undergoing unnecessary costs.

Since the models do not properly map AI agents, Habib said that maintenance includes verification prompts, model parameters, tool patterns and memory configuration. This also means tracing executions between entries, outputs, reasoning stages, tool calls and human interactions.

“You can update a [large language model] LLM’s prompt and watch the agent behave completely differently even if nothing in the history of the Git has changed, “said Habib.” The links change, the recovery indexes are updated, the APIs of the tool evolve and suddenly, the same invite does not behave only expected … It may seem that we debug ghosts. »»