Ramp moved from evaluating managed agent APIs and open‑source frameworks to delivering its own AI agent harness, Laminar, in just six weeks. The shift gives engineers direct control over model selection, debugging, session management, and credential handling—areas that directly affect reliability, cost, and security for any production AI service.
Why Build In‑House?
Ramp’s primary driver was the need to avoid lock‑in to a single model provider. By abstracting the model behind a harness, the team can switch between providers based on latency, price, or performance without rewriting agent logic. Existing observability tools were also deemed insufficient for tracing unnecessary tool calls and context bloat, prompting the addition of a built‑in debugging UI. Direct access to the harness allowed rapid iteration with product engineers, bypassing vendor release cycles.
Key Harness Design Patterns
Laminar separates three concerns:
- Execution Engine: Runs agent steps, supports asynchronous workflows so agents can continue without waiting for each sub‑task.
- Sandbox Layer: Isolates code execution from the persistent session state, allowing sandboxes to spin up and down without breaking conversation continuity. Credentials and orchestration live outside the sandbox, reducing exposure of sensitive data.
- Session Store: Persists conversation history and execution state independently of the sandbox, enabling long‑running, collaborative agents such as Ramp Inspect and Glass.
The architecture mirrors emerging patterns in other projects (e.g., LangChain’s Deep Agents), indicating a broader move toward modular, sandbox‑as‑tool designs.
Observability and Debugging Integration
Laminar embeds a debugging interface that surfaces tool invocation traces, context size, and error locations. This internal view replaces the need for external tracing pipelines and gives engineers immediate feedback when agents generate superfluous calls or encounter failures. For SRE teams, the unified observability surface simplifies alerting and root‑cause analysis, as all agent activity funnels through a single platform.
Cost‑Aware Model Routing
While the cost experiment was conducted by LangChain, the principle applies to any harness that can inspect request characteristics before dispatch. By routing simple tasks to a cheaper model and reserving high‑capacity models for complex work, organizations can achieve significant inference cost reductions without measurable performance loss. Implementing such routing at the harness level leverages the context already available about prompts, tools, and domain specifics.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Teams building production AI services should evaluate whether existing managed offerings meet their needs for model flexibility, observability, and credential isolation. If not, a lightweight internal harness—focused on sandbox separation, session persistence, and built‑in debugging—can be assembled quickly, as demonstrated by Ramp’s six‑week effort. Consider adding a routing layer to the harness to optimize inference spend, and ensure that any credential handling stays outside the execution sandbox to limit exposure. Finally, treat the harness as the operational boundary: monitor its health, instrument its internal traces, and design alerts around session failures or abnormal tool usage.
