The start-up specializing in the management of business trips and expense reports has developed its own AI agent to reserve its business trips.
Long before the “agents” became a galvanized marketing term, companies had set up functional agent worksflows. Navan is one of them. After the arrival of Chatgpt, the CTO of the start-up specializing in the management of business travel and costs quickly set out to experiment with generative models on basic reasoning. In a few months, Navan managed to create his own agentic framework named cognition using a combination of different LLM.
Use LLM to reduce hallucinations
At the time, Navan was trying to improve his customer service by automating reservations and support. In his first experiments with language models, Ilan Twig quickly came up against the limits of prompting. The instructions coded directly in the prompt proved to be extremely fragile. “It was like code inside the prompt, and this code was very sensitive,” he recalls. Each addition of functionality was likely to completely destabilize the system. The limited size of the context, initially from 4,000 tokens to 8,000 tokens, further complexed the integration of new capacities. The more tokens were used, the more the probability of errors and hallucinations increased.
To solve problems of stability and hallucination, the CTO has developed a radically different approach. Instead of managing everything by a unique prompt, he designed a multi-model system where different LLM work together, supervising each other. “I realized that to deploy something in production, you cannot do it on a single LLM,” he recalls. Based on the LLM as a judge approach, the principle is simple: a first model generates the answer to the user’s question and one or more LLM check the answer.
Cognition, a complete agent system
To go even further and fully automate the framework with external tools, the company has embarked on the development of cognition, its own agent system. The objective was to transform LLM into a functional cognitive system capable of managing complex tasks reliably and repeated. “I implemented a mechanism of thought. Take the example of a simple request like ‘Tell me what time it will be during my next trip’. Behind this apparent simplicity, the system breaks down the task into several stages: first identify the destination of the trip using a route tool, then search for geographic coordinates, consult a meteorological service, and finally formulate a precise answer”, explains Ilan Twig.
“Today we use almost 200 llm”
The system operates independently thanks to a combination of specialized tools: web search, external weather data, route tool. The model must reason on the most relevant sequence. If a tool does not respond completely, it can switch to another. “The model must understand what to do, decompose it into small components, then use the right tools,” explains the CTO. A multi-tool and multi-model approach that allows you to create a much smarter system than a simple chatbot. “Today, we use nearly 200 llm, all taking different roles, watching each other and sharing data,” says Ilan Twig.
Towards an autonomous platform
Today, the cognition power remains under-exploited for the CTO of Navan. Ilan Twig now sees its framework as an autonomous platform capable of generating applications in virtually all areas, independently. With cognition, Ilan Twig ensures that a developer is able to develop from A to Z an ERP, a new agent system or even a pizza control system, very quickly.
Ultimately, Navan’s CTO plans to present cognition as an independent platform. “Amazon has developed its own internal tools to manage its servers. And then it became AWS,” he recalls. Its objective: to create a solution allowing anyone to quickly develop intelligent agent applications. The loop will be completed.
