Ramp's Homegrown Agent Platform Signals Shift in Infrastructure Strategy

Ramp opted to construct its own AI agent harness rather than adopt an off-the-shelf solution. Following an extended evaluation period that included managed agent APIs, open-source frameworks, and third-party offerings, the company determined that building internally would deliver superior results. Laminar, the resulting platform, reached its initial release in six weeks and now underpins all agentic products across the fintech operation.

Sam Kronick, the engineer who led the project, explained the reasoning in Ramp's engineering blog: "We studied the other offerings on the market and decided that we could do better, move faster, and have stronger alignment to where we want to go by building in house." In a post published Tuesday, Ramp outlined how Laminar consolidates agent execution, sandboxing, observability, and asynchronous workflows into a unified system. The architecture drew notice from LangChain CEO Harrison Chase, who highlighted several design decisions in a post on X and referenced comparable initiatives within LangChain's Deep Agents initiative.

Though Ramp and LangChain are tackling the challenge from distinct angles, both companies point to a fundamental transformation in how agent infrastructure is being developed. As harnesses assume greater responsibilities, organizations face a critical choice regarding which infrastructure components to construct and operate internally.

Why Ramp built in-house

Ramp sought to avoid dependency on a single model provider, requiring the ability to switch between models according to cost, latency, and accuracy metrics. The company also needed deeper insight into agent behavior. Engineers were already examining execution traces to identify redundant tool invocations, excessive context, and failures, yet available observability platforms made this work unnecessarily difficult. Laminar incorporated a custom debugging interface to address this gap. Maintaining development internally also enabled the harness team to collaborate directly with product engineers rather than waiting for vendor updates.

Chase concurred with Ramp's emphasis on model independence and observability, both areas LangChain has prioritized in its own development. Additional elements of Laminar's design also captured his attention.

Sessions split from sandboxes

Laminar supports a range of applications, from customer service agents to Ramp Inspect, a cloud-based coding agent, and Glass, which manages extended, team-based tasks. This diversity of use cases demands a harness capable of maintaining persistent sessions, incorporating human approvals, and operating across multiple environments.

Chase highlighted Kronick's insight that a production harness performs substantially more than executing a basic loop. In his X post, Chase wrote "+100 to this," referencing LangChain's investment in Deep Agents.

Ramp maintains agent sessions independently from the sandboxes where code executes. Conversation records and execution state are preserved separately, permitting sandboxes to launch and terminate without disrupting the session. Chase characterizes the sandbox-as-a-tool pattern as an emerging industry trend, noting that Deep Agents employs a comparable approach. This design also maintains credentials and orchestration outside the sandbox, a critical factor as developers work to contain agents executing untrusted code.

Async agents by default

Laminar permits agents to continue operating without pausing for each step to complete. Chase describes non-blocking asynchronous agents as the coming standard and indicates LangChain is developing equivalent capabilities for Deep Agents.

Ramp's Connectors Service also drew Chase's attention. This component consolidates authentication, credentials, access controls, and tool execution behind a unified API. Chase notes that LangChain is investigating comparable functionality for managed Deep Agents. Centralizing these functions provides platform teams with a single location to govern access and credential management rather than implementing these separately for each agent.

Routing trims inference costs

The harness layer is where organizations can reduce inference expenses, a point LangChain has emphasized directly. In an October 1 engineering post, LangChain described integrating model selection into the harness of Open SWE, its open-source coding agent. A router examines the incoming request and assigns one of three models based on anticipated task difficulty, directing simpler work to GLM-5.3-Flash, intermediate tasks to GPT-5.6 Sol, and the most complex requests to GPT-6 Astra.

Across 973 agent threads, the router decreased median LLM costs from $2.61 to $0.94 per thread, representing a 64 percent reduction compared with deploying GPT-6 Astra universally. This improvement occurred without a notable decline in results. Merged pull request rates reached 29.2 percent for routed threads and 27.3 percent for the control group, a gap lacking statistical significance. Ninety percent of requests were directed to the two lower-cost model tiers.

LangChain contends that routing functions more effectively within the harness than at a generic LLM gateway, where context regarding agent operations is limited. The harness possesses knowledge of which prompts, tools, and domain context are in play, providing superior information for making that determination. LangChain employs model-selection middleware to select a model at the beginning of each thread.

The findings do not establish whether constructing a custom harness costs less than leveraging an existing framework. Ramp did not participate in the experiment. However, they demonstrate the magnitude of savings organizations can realize through improved model selection at the harness level, particularly when their agents handle tasks with disparate requirements.

The harness ownership dilemma

Ramp's choice also illustrates how coding agents have reshaped infrastructure economics. Kronick notes that agents produced all of Laminar's initial code, with seasoned engineers directing the design and execution.

Delivering the first iteration in six weeks represents one milestone. Sustaining it presents another challenge. Ramp requires engineers to manage outages, advance the platform, and assist product teams. The company has also refrained from releasing operating-cost analyses or performance comparisons against alternative platforms, making it difficult to assess whether the in-house approach delivers cost advantages.

LangChain presents an alternative path. Rather than constructing a harness from the ground up, organizations can adopt frameworks that manage much of the foundational infrastructure, a strategy other vendors are pursuing as they release their own harnesses as open source.