Ramp's Internal AI Coding Agent: Architecture and Design of 'Inspect'
This article details Ramp's decision and architectural choices behind building 'Inspect', an in-house AI coding agent. It explores why a custom solution was preferred over third-party tools, focusing on the system's ability to provide remote sandboxed development environments with deep internal integrations for verification and autonomous debugging. The design emphasizes extensibility, allowing engineers to build hundreds of specialized agents on top of the Inspect platform.
Read original on The Pragmatic EngineerThe Rationale for Building In-House AI Tools
Ramp chose to develop its own AI coding agent, Inspect, due to limitations found in existing third-party solutions. Key drivers included the need for unlimited concurrent AI agent sessions, which local machines couldn't support, and superior frontend tooling capabilities. Additionally, the increasing complexity of Ramp's systems necessitated robust remote development environments that could handle cross-system debugging and API contract verification. This decision highlights a significant architectural trade-off: build vs. buy for critical developer tooling, especially when off-the-shelf options don't meet specific operational and integration requirements.
Core Architectural Principles of Inspect
Inspect's design is founded on the principle that AI agents should operate with the same context and access to internal tools and data sources as human engineers. This allows Inspect to perform comprehensive verification (backend tests, telemetry review, feature flag queries) and visual frontend verification (screenshots, live previews) that third-party tools typically lack due to integration hurdles. This "closed-loop" verification is a crucial differentiator, enhancing the reliability and autonomy of the AI agent's output.
Key Components and Integrations
- Remote Sandboxes: Provides isolated, pre-configured development environments for each agent session, enabling high concurrency and consistent setups.
- Internal Data Access: Agents can query sanitized production database replicas, Snowflake, Looker, and dbt tables, crucial for debugging and data analysis tasks.
- Tooling Integration: Access to development services (Postgres, Redis, RabbitMQ, Temporal), Chromium for frontend work, and VS Code Server.
- Rapid Provisioning: Techniques to spin up fully provisioned remote dev environments in under 5 seconds, critical for developer productivity.
Inspect as a Platform for Agentic Workflows
Beyond its core coding and debugging functions, Inspect is designed as a platform upon which other specialized AI agents can be built. This extensibility has led to over 200 internal agents, ranging from custom code reviewers (ReviewBuddy) and on-call assistants to data analysts (Ramp Research) and automated error response systems. This platform approach demonstrates how a core system can be leveraged to empower distributed teams to develop targeted, context-aware automations, offloading the complexity of cloud backend management from individual agent builders.
Architectural Takeaway: The Power of Internal Platforms
Designing an internal platform like Inspect, which abstracts away infrastructure complexities and provides rich integration capabilities, can significantly accelerate the development and adoption of AI-powered tools across an organization. It allows domain experts to focus on agent logic rather than underlying system concerns.
Technical Stack Highlights
The Inspect platform leverages a modern cloud-native stack. Frontend interfaces are built with React/Vite. Critical backend components include Cloudflare Durable Objects for stateful serverless functionality, SQLite for lightweight data storage, and Cloudflare Agents SDK for edge computing. Modal sandboxes are used for isolated remote execution environments. This combination allows for a highly distributed, performant, and scalable architecture, capable of handling numerous concurrent agent sessions.