Ramp Opens Inspect: The Cloud Coding Agent Behind 74% of Its Merged Code

Ramp is opening Inspect, the background coding agent it built for its own engineers, to other companies. Co-founder Eric Glyman announced the launch on X on October 27, 2026, saying every product Ramp shipped this year was built with it and that "74% of the code we merge now comes from it. Last November it was close to zero." The inspect.com page currently offers early-access sign-up only.

That is a striking number, and it is also exactly the kind of figure to read carefully. This post covers what Inspect is, how Ramp has described building it, how the adoption numbers have moved, and what is claimed versus what is checkable.

Quick reference

What Inspect is

The inspect.com tagline calls it "the collaborative, cloud-based coding agent by Ramp." The page's description says it turns a task into a reviewed pull request, with shared sessions, your tools and model choice. Beyond that the page is a landing page: no feature list, pricing or documentation appears yet, so most of what is known comes from Ramp's earlier public write-ups.

The Pragmatic Engineer's piece on why Ramp built Inspect gives the motivation. Engineers liked running agents locally but could only manage one or two sessions at a time. Ramp wanted many in parallel, wanted designers to make small interface changes with AI, and wanted remote environments for debugging across systems. Most off-the-shelf harnesses lacked Ramp's internal integrations, so the team built its own.

How it works, according to Ramp's public write-ups

Modal's engineering write-up describes the design principle in one line: agents should have the same context and tools as software engineers. In practice:

• One sandbox per session. Each session runs in an isolated Modal sandbox containing Postgres, Redis, Temporal and RabbitMQ, plus a VS Code server, a web terminal and Chromium for visual checks. It connects to GitHub, Slack, Buildkite, Sentry and Datadog.
• OpenCode as the harness. Ramp chose the open-source OpenCode for its HTTP API, license and model agnosticism, per The Pragmatic Engineer.
• Snapshots for fast starts. A job runs every 30 minutes to clone repositories, install dependencies and build, then saves a filesystem snapshot; new sessions start from the latest one, so they are at most about 30 minutes behind the repository head.
• Multiplayer. Several people can feed prompts into one session through queues and locks, from Slack, a web UI or a Chrome extension that lets non-engineers select interface elements to change.
• Self-verification. The agent runs tests, checks telemetry, queries feature flags, and returns screenshots and live previews.

The same piece reports internal agents built on top of the platform: a code-review agent, an on-call incident assistant, frontend QA, data analysis, customer-feedback analysis, and automations that draft PRs from Sentry or Datadog alerts. It also reports more than 200 agents built on the platform and a core team of 5.5 people.

The adoption numbers, and why they differ

The reported share of Ramp's merged work from Inspect has moved a lot depending on date and source:

These do not line up neatly, and they should not be treated as one time series. "PRs started by Inspect", "PRs merged", and "code merged" are different measures, and different authors reported at different moments. The one clear statement is the latest: a company executive says 74% of merged code now originates from Inspect. No methodology, such as whether it counts lines, files, or pull requests, or how human edits to agent output are treated, has been published that we could find.

Timeline: from Chrome extension to product

According to The Pragmatic Engineer, Inspect's first version, about a year before that article, was a Chrome extension that let designers make small website edits. Adoption was low: engineers could already change a line of code faster themselves, and setup was too complex for non-developers. The second version, shipped in November 2025, turned it into a remote development environment with a background coding agent, and that is when adoption took off. By July 2026 the same source counts one million total sessions.

The product decision on October 27 is the third step: offering the system to other companies after a design-partner phase, as the launch coverage describes it. Ramp has also published an open-source reference for the architecture, referred to in public write-ups as Open-Inspect, which suggests the company is comfortable with the ideas being copied while it sells the polished, hosted version.

Why a fintech is selling developer tooling

Ramp is a corporate-card and spend-management company, not a developer-tools vendor, which makes the launch notable. The likely logic, which Ramp has not spelled out, is the same one behind many internal-tool spin-outs: a team that built the tool because nothing off the shelf fit its needs now has proof, in its own codebase, that it works. The 74% claim is the proof point. Whether customers outside Ramp see similar results depends on how closely their repositories, test suites and review habits resemble Ramp's, and on whether they invest in the surrounding tooling the agents depend on.

What the claim does and does not tell you

It does tell you that a well-funded engineering organization with strong internal tooling is routing most of its code changes through an agent that works in a full environment. That is consistent with a broader pattern in 2026: teams moving from single local assistants to fleets of background agents.

It does not tell you code quality, defect rates, review burden or cost. A high share of agent-authored code can coexist with a bottleneck that has simply moved to review and verification. Our coverage of the Harvard and Jellyfish analysis of AI coding agents found that more code does not automatically mean more shipped software.

How Inspect compares with other agents

Background coding agents are now a crowded category. Cognition's Devin is aimed at incident triage and engineering tasks. Local control planes such as AgentPlane run several coding CLIs side by side on one machine. Shared-session tools like agent-relay tackle handoff between agents. What distinguishes Inspect, from the public descriptions, is the combination of a full production-like environment per session, multiplayer sessions, and being proven first on the vendor's own codebase.

The isolation layer is the same problem solved by Cloudflare's Sandbox SDK and by local VM sandboxes like Capsem. If you evaluate Inspect, ask the same questions you would ask of those: what network access the sandbox has, which secrets it can read, and what is logged.

What to ask before you sign up

1. Which models can I use, and who pays for tokens? The page says model choice is supported but lists none.
2. Where does code run? Ramp's version uses Modal sandboxes; a hosted product may differ.
3. What credentials does a session hold? Agents with access to GitHub, Slack, Sentry and Datadog are powerful and need tight scoping.
4. How is quality measured? Ask for the definition behind any "percent of code" claim.
5. Pricing and data handling. Neither is published yet.

What to copy even if you never use Inspect

• Give the agent your real dev stack. Databases, queues and a browser inside the sandbox are what let it verify its own work.
• Pre-build snapshots. Starting from a recent snapshot is what makes sessions feel as fast as a local agent.
• Make it multiplayer. Letting a designer, an engineer and a reviewer steer one session beats passing a branch around.
• Choose an open, swappable harness. Ramp picked a model-agnostic one so the agent could change models.

Bottom line

Inspect is a credible, well-documented internal tool becoming a product, and the 74% figure is a real company claim by a named executive. It is also unaudited and defined loosely. Treat the numbers as an upper-bound signal of what a heavily invested team achieved, and judge the product on a pilot against your own repositories once early access opens.

Related reading

• Harvard and Jellyfish: more code, not more software
• Cognition Devin on-call: what is claimed
• AgentPlane: Claude Code, Codex and Cursor side by side
• agent-relay: shared sessions for coding agents
• Cloudflare Sandbox SDK 1.0
• Capsem VM sandbox for coding agents

Details are accurate as of October 27, 2026; early-access terms may change.