Ramp Built Its AI Agent Tool in Six Weeks

Fintech firm Ramp spent months testing AI tools, then built its own agent framework in just six weeks for better control.

Fintech company Ramp has built its own internal platform for creating and managing AI agents, a decision made after months of evaluating third-party tools. According to reporting from The New Stack, Ramp concluded that existing managed APIs, open-source frameworks, and commercial platforms were not sufficient for its needs. The resulting system, named Laminar, took a small team just six weeks to develop its first version. This custom "agent harness" now serves as the foundational infrastructure for all AI-powered agentic products at the company. Ramp's move represents a significant strategic choice to invest directly in bespoke AI infrastructure rather than relying on external providers, signaling a deliberate pursuit of greater control and performance for its core product features. This case study provides a clear example of a major technology company opting to build, rather than buy, a critical component of its AI stack when faced with the limitations of the current market offerings.

An AI agent harness, or orchestration framework, is the underlying software that enables AI agents to execute complex, multi-step tasks. It goes beyond simple API calls to a large language model by managing an agent's memory, allowing it to use various tools (like calculators or internal databases), and coordinating its actions over time to achieve a specific goal. Ramp’s decision to build Laminar suggests that off-the-shelf solutions created a "black box" problem, making it difficult to debug agent behavior, control for latency, and ensure reliability in a production environment. By building its own system, Ramp gains deep visibility into every step of an agent's reasoning process. This allows its engineers to fine-tune performance, manage costs more effectively, and rapidly iterate on new agent capabilities. A custom framework provides the granular control necessary for mission-critical applications, where predictable behavior and the ability to quickly diagnose failures are non-negotiable requirements that generic platforms often struggle to meet.

Ramp's journey is part of a broader industry shift, as noted by LangChain CEO Harrison Chase. Many organizations initially adopted high-level frameworks and managed APIs to quickly prototype AI features. This "buy" approach prioritizes speed and lowers the barrier to entry for experimentation. However, as companies move from these initial proofs-of-concept to deploying robust, scalable products, they often encounter the inherent limitations of generalized tools. The need for specialized logic, optimized performance, and deep integration with proprietary data and systems becomes a critical bottleneck. This evolution mirrors previous cycles in software infrastructure, such as the transition many companies made from simple platform-as-a-service (PaaS) solutions to more complex but powerful systems like Kubernetes. As the AI stack matures, the most sophisticated teams are realizing that a one-size-fits-all approach to agent orchestration is insufficient for building a lasting competitive advantage, pushing them to bring this core infrastructure in-house.

For CTOs, founders, and engineering leaders, Ramp's experience reframes the "build vs. buy" debate for AI infrastructure. The central question is no longer just about choosing a foundational model, but about deciding who controls the orchestration layer that brings that model to life. For simple, non-critical applications, leveraging third-party APIs remains a sensible and efficient strategy. However, for complex AI agents that are central to a core product offering, building a custom framework may be the only way to achieve the desired level of performance, reliability, and user experience. Teams should evaluate their needs based on factors like debugging complexity, latency sensitivity, and the need for custom tool integrations. Looking ahead, the industry should expect to see more companies follow Ramp's lead. This will likely spur the development of a new class of open-source and commercial tools that offer a middle ground: providing more control and transparency than fully managed APIs, but without the heavy lifting of a completely bespoke build from the ground up.

Ramp's decision to build a custom AI agent harness demonstrates that off-the-shelf frameworks may not meet the performance and control demands of production systems. For engineering leaders, this shifts the 'build vs. buy' calculation, suggesting that bespoke infrastructure is becoming a viable, even necessary, path for deploying complex agents.

Building custom AI infrastructure like Ramp's can create a significant competitive moat, enabling unique product features that off-the-shelf tools can't replicate. However, this strategy requires substantial upfront investment and specialized talent, representing a major resource commitment that could divert focus from other core business activities.