Ramp built its own coding agent, Inspect, and The Pragmatic Engineer's deep dive on Ramp's Inspect reported on August 25, 2026 that 75 percent of the company's merged pull requests come from it. If that story has your team asking whether to build your own, here are the four questions we'd run first.
We run production agentic delivery pipelines on bought harnesses (a harness is the software that runs the model's agent loop and its tools), with our own verification and approval layer on top. So we've had to answer this one for ourselves. For most mid-market teams, the answer is to buy the agent and own a thin layer around it, unless you can name a constraint no vendor will fix.
1. Can you name the constraint a vendor won't fix? Ramp's reasons were specific. Local machines capped how many agents could run at once. Frontend work needed better tooling. The team needed remote dev environments as complexity grew. Those are infrastructure problems with a shape. "We want more control" doesn't have one. Write your constraint down in one sentence, then ask a vendor whether they solve it. Sometimes they already do.
2. Can a bought agent reach and verify against your context? Inspect's edge isn't a secret model. It's plumbing. The agent checks backend changes against tests, telemetry, and feature flags, and checks frontend work against screenshots and live previews. Steal that question. If an off-the-shelf agent with standard connectors can see what it needs to verify its own work, the build case just lost its best argument. If that context sits behind internal systems nothing can reach, now you're talking.
3. Are you buying a tool or building a platform? Ramp's engineers have built more than 200 agents on top of Inspect. That's a platform, and platforms compound. If what you actually want is pull requests, you're describing a tool, and you should buy tools. Be wary of a build case that only pays off once the platform future shows up.
4. Who owns it in year two? The Inspect team is 5.5 people: four engineers, a director, and a part-time PM. More than 150 Ramp engineers have contributed to the codebase. A homebrew harness is a product. It has users, an upgrade treadmill, and an on-call rotation, and it has to keep pace with every model release. If that headcount made you flinch, you have your answer. It's the answer most mid-market teams should hear.
Notice what Ramp didn't build: the agent loop itself. Inspect runs on OpenCode, an open-source harness they picked partly because it's model-agnostic. That's the lesson on the buy side too. Whichever way you go, the parts you own should survive a switch of model or vendor. We covered the portability checks we run before committing to an agent separately. Build versus buy is the question you answer before those checks.
That thin layer you own (verification hooks, approval gates, containment) is where the durable advantage tends to live, and it's small enough to maintain. If you want a second opinion before you commit headcount, our AI implementation framework is where we'd start.