FinTech enterprise Ramp has made a decisive move in the artificial intelligence infrastructure landscape by developing and deploying an internal agent harness named Laminar, bypassing third-party platforms and commercial managed agent APIs. Following months of exhaustive testing involving various open-source frameworks, managed offerings, and third-party solutions, Ramp’s engineering team concluded that an in-house platform would deliver superior performance, faster iteration cycles, and closer strategic alignment with corporate objectives. Conceived and engineered primarily by Ramp software engineer Sam Kronick, the foundational version of Laminar was built in a remarkably brief window of six weeks. Today, the proprietary harness serves as the central operational backbone for all agentic products deployed across the fintech firm, representing a growing trend among sophisticated technology enterprises seeking deep infrastructural control over their autonomous systems.
The Genesis of Laminar: A Chronology of Evaluation and Development
The timeline leading to the launch of Laminar began with Ramp’s ambitious push into agentic artificial intelligence applications aimed at automating internal workflows, customer support, and software engineering processes. Initially, the fintech enterprise experimented with a wide array of externally available managed agent APIs and community-driven open-source frameworks. However, engineers quickly encountered architectural bottlenecks, including vendor lock-in risks, opaque error logs, and difficulties in tailoring external models to specific business needs.
Realizing that generic tools lacked the nuance required for high-stakes financial operations, the engineering team pivoted toward internal development. In late 2024 and early 2025, Sam Kronick spearheaded the initiative to design a unified execution environment. In a notable demonstration of modern software development speed—and a testament to the utility of AI-assisted coding—AI agents themselves reportedly generated the vast majority of Laminar’s initial codebase, while experienced engineers provided high-level architectural guidance and rigorous oversight. Within just six weeks of conception, the first stable version of Laminar was operational, quickly scaling to handle critical production workloads such as Ramp Inspect, a cloud coding agent, and Glass, a specialized system designed for long-running, collaborative background tasks.
Architectural Innovations: Separating Sessions from Sandboxes
In a technical blog post published on Ramp’s engineering blog in Tuesday, the company detailed how Laminar successfully consolidates agent execution, secure sandboxing, system observability, and asynchronous workflows into a single cohesive platform. Among its most praised architectural choices is the deliberate separation of agent sessions from the sandboxing environments where untrusted code is executed.
In traditional agent architectures, conversational history, execution states, and code execution environments are often tightly coupled, creating significant vulnerabilities and scalability bottlenecks. Laminar solves this by decoupling session state and conversation history from the underlying sandboxes. This structural division allows individual sandboxes to spin up, execute tasks, and shut down dynamically without disrupting the overarching session or losing contextual data. Furthermore, this design isolates credentials, access controls, and orchestration layers outside the vulnerable execution zone—a critical security consideration as enterprises increasingly deploy agents capable of running untrusted or dynamically generated code.
Observability, Asynchronous Workflows, and the Connectors Service
Beyond secure sandboxing, Laminar addresses a notorious pain point in modern AI engineering: observability. Prior to building Laminar, Ramp engineers spent considerable time manually parsing complex execution traces to identify redundant tool calls, bloated context windows, and execution errors. Existing third-party observability tools often proved cumbersome or insufficiently granular. To resolve this, Ramp integrated a custom debugging interface directly into Laminar, offering engineers instantaneous visibility into agent behaviors.
The platform is also designed for asynchronous operation by default. Unlike synchronous systems that stall while waiting for lengthy tasks to conclude, Laminar enables agents to persist with complex workflows autonomously, seamlessly incorporating human-in-the-loop approval gates when necessary. To manage external tool integration securely, Ramp developed the Connectors Service. This centralized microservice standardizes authentication, credential management, access control lists, and tool execution behind a unified API, giving platform administrators a single point of governance rather than decentralized, ad-hoc implementations across individual agents.
Industry Reaction: LangChain and the Broader Paradigm Shift
Ramp’s engineering decisions have reverberated across the broader software development community, drawing public commentary from prominent industry leaders. Harrison Chase, CEO and co-founder of LangChain, highlighted Laminar’s design philosophy on social media, drawing direct parallels to ongoing developments within LangChain’s own open-source "Deep Agents" initiative.
Chase expressed strong agreement with Ramp’s prioritization of model independence and deep observability, emphasizing that production-grade agent harnesses must evolve far beyond simple execution loops. While Ramp and LangChain are tackling the complexities of agent infrastructure from slightly different angles—Ramp through bespoke in-house engineering and LangChain through comprehensive open-source frameworks—their concurrent trajectories signal an undeniable industry shift. Enterprises are increasingly recognizing that the foundational harness layer dictates the ultimate success, security, and cost-efficiency of autonomous AI applications.
The Economics of Inference: Harness-Level Model Routing
One of the most compelling aspects of modern agent infrastructure is its capacity to optimize large language model (LLM) inference costs. Because inference expenses scale rapidly with complex agentic loops involving thousands of reasoning steps, finding ways to route requests efficiently is paramount. LangChain recently published data regarding its own harness-level model router implemented in Open SWE, its open-source software engineering agent. By evaluating incoming requests against task complexity at the harness level rather than relying on generic LLM gateways, LangChain demonstrated dramatic cost reductions.
The model router dynamically assigns tasks to one of three tiers based on anticipated complexity: routing simpler operations to cost-effective models like GLM-5.3-Flash, intermediate tasks to GPT-5.6 Sol, and highly demanding architectural reasoning to flagship models like GPT-6 Astra. Across nearly one thousand agent threads, this routing strategy reduced median LLM costs from $2.61 per thread to just $0.94 per thread—a striking 64 percent decrease. Crucially, these savings were achieved with zero statistically significant impact on task performance, as measured by successful pull request merge rates (29.2 percent for routed threads versus 27.3 percent for the control group). Approximately 90 percent of all requests were successfully handled by the two more economical model tiers.
Industry analysts note that while Ramp’s specific cost metrics for Laminar remain proprietary, the underlying economic logic aligns with industry-wide findings. Routing models intelligently inside the harness leverages deep contextual awareness of the agent’s immediate goals, prompts, and toolsets—information that generic, stateless gateways simply do not possess.
The Strategic Dilemma: Build In-House Versus Adopt Frameworks
Ramp’s successful deployment of Laminar highlights a central strategic dilemma facing modern technology organizations: whether to build and maintain custom agent harnesses in-house or adopt established third-party frameworks.
Building a custom platform grants an organization absolute control over its technology stack, enabling rapid feature delivery tailored precisely to internal product roadmaps and eliminating dependencies on external vendors. Furthermore, as demonstrated by Ramp’s rapid six-week development cycle aided by AI coding assistants, the initial barrier to creating bespoke software has lowered significantly.
However, industry experts caution that initial development speed does not equate to long-term operational ease. Maintaining an internal harness requires dedicated engineering resources to handle security patches, framework updates, incident response, and continuous optimization as underlying model capabilities evolve. Without published benchmarks detailing total cost of ownership—including ongoing maintenance and infrastructure overhead—smaller enterprises may find the in-house route prohibitively resource-intensive.
Conversely, open-source frameworks and managed platforms offered by organizations like LangChain and major cloud providers provide a viable alternative for companies lacking the specialized engineering bandwidth to construct and maintain custom harnesses from scratch. As the ecosystem matures and more vendors open-source their specialized agent execution environments, technology leaders must carefully weigh the strategic advantages of infrastructural autonomy against the operational burdens of long-term software maintenance.
