Ramp's Inspect: An Internal Coding Agent That Powers 30% of Engineering Pull Requests

Ramp's engineering team has shared the architecture of Inspect, an internal coding agent that has achieved approximately 30% adoption for merged pull requests across the company's frontend and backend repositories. This represents a significant milestone in AI-assisted software development, demonstrating how deeply integrated, purpose-built agents can achieve substantial adoption without mandates.

The Core Innovation: Full Development Environment Access

What sets Inspect apart from typical AI coding assistants is its comprehensive access to Ramp's engineering ecosystem. Instead of generating code in isolation, Inspect operates within the same sandboxed virtual machines on Modal that human engineers use. This allows the agent to interact with databases, CI/CD pipelines, monitoring tools like Sentry and Datadog, feature flag systems, and communication platforms including Slack and GitHub.

The agent doesn't just write code—it validates it through the same testing and monitoring processes engineers rely on daily. This creates a closed verification loop that addresses what Ramp identifies as the "verification gap" affecting most AI coding tools. Inspect can execute tests, query monitoring dashboards, check database states, and participate in code reviews, ensuring generated code meets production standards.

Technical Architecture: Modal and Cloudflare Durable Objects

Ramp's choice of Modal as the execution platform proves central to Inspect's performance. Modal enables near-instant session startup and supports unlimited concurrent sessions, allowing multiple engineers to work with separate agent instances simultaneously without resource contention. The platform's sandboxing capabilities and file system snapshots maintain security while enabling rapid iteration cycles.

For state management, Ramp implemented Cloudflare Durable Objects to maintain conversation context and development session state across interactions. This stateful design allows agents to retain awareness of their work, similar to how human engineers remember codebase context during development. The architecture ensures continuity when agents switch between tasks or when engineers resume conversations with the agent.

Multi-Modal Interaction Patterns

Recognizing that different development tasks benefit from different interaction patterns, Ramp built multiple client interfaces for Inspect:

• Slack bot for quick queries and simple tasks
• Web interface for complex, multi-step development workflows
• Chrome extension specifically designed for editing visual React components

This multi-modal approach acknowledges that a single interface cannot optimally serve all development scenarios. The system also supports collaborative workflows, allowing team members to observe and guide agent actions in real-time. This addresses a common concern about autonomous coding tools by maintaining human oversight while capturing automation benefits.

Build vs. Buy: The Integration Advantage

Ramp makes a deliberate case for building rather than purchasing off-the-shelf coding agent solutions. The engineering team argues that internal tooling enables significantly deeper integration than commercial products, particularly for connecting with proprietary systems, databases, and workflows that external vendors cannot access.

This approach requires substantial engineering investment, but Ramp believes the competitive advantage comes from execution rather than architectural secrecy. The company has shared detailed implementation specifications, including execution environments, agent integration patterns, state management approaches, and client implementation details.

Organic Adoption and Productivity Impact

Perhaps most striking is that Inspect's 30% adoption rate occurred without any mandate. Engineers chose to use the agent because they found tasks where it matched or exceeded manual coding in quality, speed, or convenience. The continued growth trajectory suggests expanding comfort with the system's capabilities and limitations.

Inspect also lowers the barrier to code contribution, potentially enabling non-engineers like product managers and designers to add code directly using the same tools as professional developers. This could reshape how cross-functional teams collaborate on software projects.

Limitations and Considerations

Ramp's engineering team acknowledges that session speed and quality remain constrained by current language model intelligence. Despite sophisticated tooling and infrastructure, coding agents still struggle with complex reasoning, hallucinations, and require human oversight for critical decisions.

The company also notes that their build-versus-buy recommendation may not apply to every organization. Implementing a similar system requires strong AI infrastructure expertise and engineering resources that smaller teams or different organizations might lack or find difficult to justify.

Broader Implications for Engineering Teams

Inspect demonstrates that with appropriate context, tools, and verification mechanisms, AI coding agents can substantially boost engineering productivity at scale. The architecture provides a blueprint for organizations considering similar implementations, showing how to bridge the gap between code generation and production-ready software.

As coding agents continue evolving, Ramp's technical specifications and adoption metrics offer concrete data points for organizations assessing their automation strategies. The success of Inspect suggests that the future of AI-assisted development lies not in generic code generation, but in deeply integrated, context-aware systems that understand and operate within specific engineering ecosystems.

For teams considering similar approaches, Ramp's experience indicates that the key differentiator isn't just the AI model's capability, but the surrounding infrastructure that provides context, verification, and seamless integration with existing development workflows.

---

Related Resources:

• Modal Platform - Serverless compute platform for AI and ML workloads
• Cloudflare Durable Objects - Stateful serverless compute
• Sentry - Application monitoring and error tracking
• Datadog - Cloud-scale monitoring and analytics