Leonardo Felipe Nerone
Leonardo Felipe Nerone

An AI agent can sound capable and still be almost useless inside a company. It may answer general questions well, summarize a document cleanly, or produce confident output in a narrow test. Then it meets the real business environment: scattered data, permission limits, old systems, sensitive information, unclear workflows, and decisions that require more than fluent language. For Leonardo Felipe Nerone, co-founder and CTO of Drachma, that is where the future of AI will be decided.

"A better model does not fix a system that has the wrong context," Nerone says. "If the agent does not know what it should know, cannot access what it needs, or sees information in the wrong way, the result will still be weak."

Drachma builds custom AI systems for businesses, and Nerone's work has pushed him toward a specific question: what does an AI system need to know, access, remember, and ignore in order to be useful? That question sits at the center of context engineering, a field Nerone believes will separate serious AI systems from expensive experiments.

Context engineering is the practice of giving an AI system the right information, tools, memory, permissions, and business understanding at the right time. It moves the conversation beyond prompt engineering and into the deeper structure of how an AI system operates.

"Context is not decoration around the model," Nerone says. "It is part of the architecture."

That is important as companies move from simple AI assistants to agents that can plan, retrieve information, use tools, and complete multi-step workflows. A chatbot that answers basic questions may tolerate shallow context. An agent that touches internal systems cannot. Once AI starts operating inside business workflows, the quality of its context affects accuracy, reliability, security, cost, and trust.

Nerone explored this idea in his Forbes Council article "Context Engineering Is Cost Engineering," where he argued that memory and context are not only technical features. They are part of the economic structure of scalable AI systems.

"Every decision about context has a cost," he says. "What the agent sees, how long it thinks, what it retrieves, what it stores, and which tools it calls all affect performance and economics."

That is one reason Nerone believes many AI discussions are too model-centered. Companies often ask which model is best, which tool to buy, or how autonomous an agent can become. Those questions matter, but they are not enough. The stronger question is how the entire system is designed around the model.

If an AI system receives too little context, it may produce shallow or incorrect answers. If it receives too much, it may become slower, more expensive, and harder to control. If its memory is poorly designed, it may preserve information that is no longer useful or miss information that matters. If permissions are too loose, it may create risk. If permissions are too restrictive, it may fail to help.

"The hard part is not just adding more context," Nerone says. "The hard part is giving the system the right context for the right task, without making it slower, riskier, or more expensive than it needs to be."

This is where Drachma's work becomes more specific than selling a general AI layer. The company builds around each client's actual operation: databases, APIs, internal tools, business rules, security requirements, user needs, and workflow design. For one client, Drachma built a custom chatbot connected to the client's data environment, allowing users to interact with complex data in natural language instead of relying on manual analysis or technical navigation.

The broader lesson is that context cannot be treated as an afterthought. It has to be designed into the system from the beginning.

"A company's data is not just a pile of information," Nerone says. "It has structure, permissions, history, and meaning. AI has to respect that or it will not be trusted."

Trust is a recurring theme in Nerone's view of AI agents. Many companies want faster workflows and more automation, but they also need control. They care about logs, human review, data boundaries, access rights, and reliability. For Nerone, autonomy is only useful when the context is strong enough to support it.

"A system that acts without the right boundaries, permissions, and memory does not become more valuable because it moves faster," he says. "It becomes harder to trust."

An AI assistant that helps an employee find an internal policy may need one kind of access. An agent that analyzes customer data may need another. A system embedded into a client's product may need stricter boundaries, clearer evaluation, and more careful design around failure.

Nerone's technical background spans payments, fintech infrastructure, crypto, open-source development, data systems, and enterprise software. That range gives him a practical view of what happens when software has to work across systems, users, and constraints.

"People see the answer on the screen," he says. "They do not always see the retrieval, the permissions, the memory, the data cleaning, the evaluation, or the cost trade-offs behind it."

That hidden architecture is becoming more important as businesses move beyond individual productivity tools. A simple AI tool can help one person write faster or summarize a file. A company-level AI system has to do more. It must understand who is asking, what they are allowed to see, where the relevant information lives, how current that information is, and whether the answer can be trusted.

Context engineering also forces companies to think about cost with more discipline. More retrieval, more memory, longer reasoning, and broader tool access may improve output in some cases, but they can also increase latency and expense. The best system is not always the one that uses the most context. It is the one that uses the right context efficiently.

"That trade-off is central," Nerone says. "There is a relationship between accuracy, time, and cost. Context engineering is how you manage that relationship instead of pretending it does not exist."

Drachma's current focus reflects that mindset. Nerone wants the company to become a lean, technically excellent business serving clients in the United States and Brazil, with reusable technical foundations that make custom AI systems faster and more reliable to deliver. He sees context engineering as central to that future because it determines whether AI can move from impressive output to dependable operation.

The winners in AI, he believes, will not simply have better prompts or larger models. They will know how to build the environment in which AI can work.

"The companies that win will not just ask better questions," Nerone says. "They will build better systems around the model."

That is the practical edge Nerone is pursuing at Drachma. AI agents may be one of the most discussed ideas in technology, but their value will depend on less visible design choices: what they remember, what they retrieve, what they are allowed to touch, how they are evaluated, and what it costs every time they run.

For Nerone, that is where the real technical race is happening.

"Context engineering is where AI becomes useful or expensive noise," he says. "If you get it right, the system becomes more reliable, more efficient, and more connected to the business. If you get it wrong, the model will not save you."

For information on Leonardo Felipe Nerone, visit his LinkedIn.