Imagine a world where 30% of your engineering work is done automatically by an AI. That's not science fiction; it's the reality Ramp, the fintech company, is living in today. But here's the real kicker: they built this AI in-house.
Ramp has unveiled (and meticulously documented at https://builders.ramp.com/post/why-we-built-our-background-agent) the architecture of their internal coding agent, named Inspect. This isn't just some fancy code suggestion tool; Inspect is a full-fledged member of the engineering team, contributing to roughly 30% of merged pull requests across both their frontend and backend codebases. The secret? Ramp gave Inspect the keys to the kingdom.
What makes Inspect truly revolutionary is its complete access to Ramp's entire engineering ecosystem. Forget limited AI assistants that only spit out snippets of code. Ramp's system, running in secure, sandboxed virtual machines powered by Modal, can interact with everything a human engineer can. We're talking databases, CI/CD pipelines, monitoring tools like Sentry and Datadog, feature flags, and even communication platforms like Slack and GitHub. Think of it as a junior engineer who never sleeps, constantly learning and contributing.
Ramp's engineers emphasize that this comprehensive verification loop is a game-changer compared to older, simpler code generation tools. Inspect can run tests, monitor dashboards, query databases for validation, and participate in code reviews. This effectively closes what they call the "verification gap" – a common problem where AI coding assistants generate code that isn't properly tested or validated. It's like giving the AI a quality assurance team built-in.
Ramp's strategic decision to leverage Modal's infrastructure is crucial to Inspect's performance. Modal enables near-instant session startups and supports unlimited concurrent sessions. This allows multiple engineers to work with separate instances of Inspect simultaneously without resource contention. Furthermore, Modal's robust sandboxing and file system snapshots ensure code execution safety and rapid iteration cycles. It allows for a secure and scalable environment for the AI to operate within.
The architecture also incorporates Cloudflare Durable Objects for state management. This ensures that the conversation context and development session state are maintained across interactions. This stateful design allows Inspect to "remember" what it's working on, much like a human engineer keeps the codebase in mind during development. This is a critical element that allows for more complex and contextually relevant code generation.
To make Inspect accessible across various workflows, Ramp implemented multiple client interfaces. Engineers can interact with the agent via a Slack bot for quick queries, a web interface for detailed tasks, and even a Chrome extension specifically designed for editing visual React components. This multi-modal approach acknowledges that different tasks require different interaction styles, making the agent adaptable to the user's needs.
The system also fosters collaboration. Team members can simultaneously observe and guide the agent's actions. This addresses a common concern about fully autonomous coding tools: the lack of human oversight. By keeping humans in the loop, Ramp ensures that the benefits of automation are realized without sacrificing control or quality.
Ramp explicitly advocates for building internal coding agent solutions rather than purchasing off-the-shelf products. The core argument is that owning the tooling enables far deeper integration with proprietary systems, databases, and workflows that external vendors simply cannot access. This control allows for tailored solutions that perfectly fit the company's specific needs.
Of course, Ramp acknowledges that this approach requires significant engineering investment. To inspire others, they've openly shared detailed implementation specifications, including execution environments, agent integration patterns, state management, and client implementation details. This transparency suggests a belief that competitive advantage lies in execution, not in hiding architectural secrets.
But here's where it gets controversial... Ramp didn't force anyone to use Inspect. The impressive 30% adoption rate happened organically, as engineers voluntarily embraced the agent. This suggests that Inspect consistently delivers code of comparable quality, speed, and convenience to manual coding. The continued upward trend points to a growing comfort level with the system's capabilities and limitations.
The team also highlights that Inspect democratizes code contribution by providing non-engineers with access to the same tools used by professional developers. This could potentially empower product managers, designers, and other stakeholders to directly contribute code, fundamentally altering cross-functional collaboration. Could this lead to a future where everyone codes?
Ramp's engineering team is realistic about the limitations. They recognize that session speed and quality are ultimately constrained by the intelligence of the underlying language models. Even with optimal tools and infrastructure, coding agents are still prone to errors, hallucinations, and struggles with complex reasoning, requiring human oversight.
And this is the part most people miss... Ramp acknowledges that their build-versus-buy recommendation might not be universally applicable. Successfully implementing a similar system necessitates robust AI infrastructure skills and substantial engineering resources. Smaller teams or organizations with different priorities might lack these resources or deem them unjustifiable.
As coding agents continue to evolve, Ramp's technical specifications and adoption metrics provide valuable data points for organizations evaluating their automation strategies. Inspect demonstrates that, with the right context, tools, and verification mechanisms, AI coding agents can significantly enhance engineering productivity on a large scale.
So, what do you think? Is Ramp's approach the future of software development, or is it a solution tailored to a specific set of circumstances? Will we see a surge in internal AI coding agent development, or will companies primarily rely on off-the-shelf solutions? Share your thoughts in the comments below!
