Why enterprises are building AI coding layers instead of buying them—and still paying for the models

A clear architectural pattern has emerged across major tech companies building AI-assisted development tools: the strategic value lies not in creating their own reasoning models, but in controlling the software layer that orchestrates them. Coinbase, Shopify, and Ramp each independently developed internal coding agents—yet all three continue purchasing access to models like Claude, Codex, and other frontier offerings.

This points to something more nuanced than a standard build-versus-buy choice. Companies are selectively insourcing the agent harness—the execution framework that manages context, security, tool access, and workflow automation—while outsourcing the underlying language models to specialized providers.

A repeatable pattern emerges#

The three companies' independently developed systems—Coinbase's Forge, Shopify's River, and Ramp's Inspect—tackle nearly identical problems. Forge connects Slack, GitHub, and Linear to automate the journey from bug report to pull request. River operates within Shopify's centralized engineering environment and now touches roughly one in eight merged code changes. Inspect runs inside sandboxed virtual machines and integrates with observability and deployment tools like Datadog and Buildkite.

Despite their different origins, all three converge on the same principles: sandboxed execution spaces, secure system integration, workflow orchestration, and automated quality checks before code ships.

LangChain's recent open-source release of Open SWE validates that this has become an industry-standard architecture, with the framework matching patterns already running inside Stripe, Ramp, and Coinbase.

Owning the gateway, not the engine#

The competitive advantage for these enterprises sits in their control over the intermediary layer. By owning the platform that sits between developers and foundation models, companies can implement central governance that individual engineers cannot.

Coinbase exemplifies this shift. The company now runs more tokens through its infrastructure while cutting overall spending through architectural choices: intelligent request routing, lower-cost default models as fallbacks, prompt caching, and tighter visibility into AI expenditure. The infrastructure team tested open-weight alternatives like GLM 5.2 and experimented with model selection without requiring thousands of engineers to switch tools.

Rather than asking developers to adopt new AI services, platform teams push updates to routing logic. Model upgrades and pricing experiments become infrastructure operations.