Ramp built an AI agent harness in six weeks — and LangChain’s CEO says it’s part of a bigger shift

Ramp, a fintech company, built its own AI agent harness called Laminar in six weeks after testing various managed agent APIs, open-source frameworks, and third-party platforms. The decision to build in-house was driven by the need for flexibility, faster iteration, and stronger alignment with Ramp's goals. Sam Kronick, an engineer at Ramp, explains that the harness brings together agent execution, sandboxing, observability, and asynchronous workflows under a single platform.

LangChain CEO Harrison Chase praised Ramp's architecture, highlighting its design choices and mentioning similar work in LangChain's Deep Agents. Both Ramp and LangChain are taking different approaches to agent infrastructure, but they are both pointing to a bigger shift in the field. As harnesses take on more responsibility, companies must decide how much of that infrastructure they want to build and maintain themselves.

One of the key advantages of building an in-house harness is the ability to avoid being locked into a single model provider. Ramp needed the flexibility to switch models based on cost, speed, and performance, as well as better visibility into how its agents behaved. Engineers at Ramp were already debugging unnecessary tool calls, bloated context, and errors, but existing observability tools made this process more challenging. By building their own debugging interface into Laminar, Ramp addressed these issues.

Another benefit of building in-house is the ability to work directly with product engineers instead of waiting on vendors to make changes. This direct collaboration enables the harness team to better understand product requirements and implement changes more efficiently.

Ramp runs various products, including customer service agents, a cloud coding agent, and a collaborative work agent, using Laminar. To support these products, the harness must manage persistent sessions, human approvals, and work across different environments. Laminar keeps agent sessions separate from the sandboxes where code runs, allowing conversations and execution states to be stored independently.

This separation enables sandboxes to spin up and shut down without disrupting ongoing sessions, providing greater flexibility and stability.

LangChain's CEO, Harrison Chase, noted that a production harness does far more than run a simple execution loop. LangChain is investing in Deep Agents and has highlighted Kronick's observation in his X post. As companies adopt harnesses, they will need to manage more than just running agent code.

An important consideration for building a harness is handling asynchronous agents by default. Laminar allows agents to keep working without waiting for every step to finish, while LangChain is also developing similar functionality for Deep Agents. This non-blocking approach is seen as the future of agent infrastructure.

Centralizing functions within the harness, such as authentication, credentials, access controls, and tool execution, can also lead to cost savings. Ramp's Connectors Service centralizes these functions behind a shared API, making it easier for platform teams to manage access and credential handling. LangChain has explored similar features for managed Deep Agents, and the results show significant inference cost reductions.

By assigning tasks to different model tiers based on complexity, LangChain managed to cut median LLM costs from $2.61 to $0.94 per thread, a 64% drop compared to using a single high-cost model for every task.