The coding agent behind 60% of Ramp’s merged PRs
Ramp adopted Linear when the company was just five people. Over 1,500 employees later, with their entire product workflow running through Linear, Zach Bruggeman (an engineer at Ramp) wanted to see what else it could power. He’d been watching AI agents grow more capable, and along with a couple of his colleagues, spent two weeks building a background coding agent tightly integrated into the workflows Ramp already ran through Linear.
Businesses shuttle many billions of dollars along the lines of code in Ramp’s codebase each year. Yet today, over 60 percent of all merged PRs at the company have been authored by Inspect, Ramp’s internal coding agent; demonstrating how an organization operating in a sensitive regulatory environment can scale AI adoption responsibly.
How Ramp built a coding agent for its own stack
Inspect is a background coding agent Ramp built themselves. It writes code and verifies the output using the same context and tools a Ramp engineer would. For backend tasks, Inspect can run tests, review telemetry, and query feature flags; and for frontend work, it can visually verify changes and return screenshots or live previews. Each session runs in a sandboxed environment configured with the infrastructure Ramp engineers already use.
Inspect began as a casual idea shared by Zach and two of his colleagues, Jason Quense and Rahul Sengottuvelu. Instead of defaulting to one of the many coding agents already on the market, they decided to build an agent optimized for Ramp’s internal workflows. “It’s a really tight integration with our development lifecycle and our tooling,” Zach says.
Being an internal tool, the team could be precise about how the agent connected to Ramp’s other systems. As Zach put it, “we know our one API key for this” and “we know exactly what the schema of these logs is going to be.” This helped them build the agent faster, while also improving its performance because there are fewer layers of abstraction between Inspect and the company’s tooling.
One of the most important of those tools is Linear. Ramp runs its entire product development process through Linear, which means the platform holds a structured record of everything associated with building product at the company. The specs teams write, the customer feedback they collect, roadmaps they plan against, and the relationships between all of it.
By integrating Inspect through Linear’s API for agents, the Ramp team gave the agent native access to that entire layer of context. And because Inspect can also reach into Ramp’s codebase, internal documentation, and other systems, the combination means the agent operates with something close to the full picture a well-informed engineer would have. When someone asks Inspect to fix a bug or think through a feature idea, it can trace the thread from a customer request through a product spec to the relevant code, rather than working from a narrow prompt.
Zach described the connection as a “natural fit” noting that the API was especially well-suited to background agent workflows and has become one of the more popular integrations internally. Building the integration was also fast. Zach says he prompted a coding agent with Linear’s docs, added a couple of rough pointers, and it got the integration “90 percent of the way there in about 30 minutes.”
How Inspect scaled AI workflows at Ramp
Inspect gained traction inside Ramp through multiple entry points, including a web interface, a Chrome extension, and, most potently, a Slack integration. Slack brought Inspect out into the open, and when people saw the agent in action, they were more likely to try it themselves; a pattern we tracked when Linear’s Slack integration was introduced.
The visibility also helps engage non-engineering teams like Design, Product Management, Product Ops, and Data. Zach’s premise is that people in these roles already have the context to guide good product decisions, and Inspect allows them to apply that expertise without needing to start by going through Engineering.
For example, while reworking a dashboard panel on Ramp’s web app, a designer might start in a Figma file and use Inspect to get the change most of the way there, before linking the session to an engineer. As a next step, instead of bothering to download code locally, Engineering can easily step in and bring the redesign live.
Inspect was built to be “multiplayer” to support exactly this kind of handoff. If someone starts a session with the agent, they can share a link to it with their teammate for further input. Alternatively, if Inspect was being directed from Slack, others can jump into the same thread and continue the work there. That shared workflow is becoming one of Inspect’s most prominent patterns of use. The result is that non-technical teams can carry ideas further into production on their own, leaving engineers to review work that’s already most of the way there.
This model works in part because Ramp built shared rails for AI adoption. A central team defines how agents can access internal systems and what they’re allowed to do, which makes it easier (and safer) for functional teams to build their own custom automations safely.
Software development’s next constraint
Inspect has accelerated Ramp’s culture of acting promptly on product feedback, instead of letting small issues languish in the backlog. As Zach describes it, feedback comes in through Slack, someone replies with “@Inspect, fix this,” a PR opens, and once approved, the change ships.
That speed has shifted the bottleneck from writing code to reviewing it. A new challenge any company using coding agents at scale will face. Ramp is exploring ways new tooling can help them overcome this constraint and move even faster. Zach says he’s paying close attention to how companies like Linear are shaping the review process.