Enterprise engineering teams are converging on a common architecture for AI-assisted software development. Coinbase, Shopify, and Ramp have each built an internal coding agent for their developers. Yet none of those house-built agents have replaced commercial tools such as Claude Code, Codex, or Cursor.That makes this less a build-versus-buy decision and more an architectural one.The layer these companies chose to own is not the large language model. It is the agent harness – the execution environment that provides context, permissions, workflow orchestration, tool access, and verification. Frontier models from Anthropic, OpenAI, and Google continue to provide the reasoning engine, while enterprises increasingly own everything around it.Frontier models from Anthropic, OpenAI, and Google continue to provide the reasoning engine, while enterprises increasingly own everything around it.LangChain recently acknowledged that this architecture has become repeatable by open-sourcing Open SWE, describing it as the public implementation of the same pattern already deployed internally at Stripe, Ramp, and Coinbase.The competitive advantage sits outside the modelAlthough Forge, River, and Inspect were developed independently, they solve remarkably similar problems.Coinbase’s Forge integrates Slack, GitHub, and Linear into a workflow that can convert a bug report into a pull request with minimal developer intervention. Once an issue is identified, Forge retrieves the required repository context, generates code, validates the changes, and even produces a mobile build for review.Shopify’s River follows the same architectural approach. The company attributes much of its success not to the agent itself but to the underlying developer platform. River operates on top of a reproducible engineering environment backed by a company-wide monorepo, durable execution sessions, sandboxed environments, credential management, and reusable engineering skills. Shopify reports that River now participates in one out of every eight merged pull requests.Shopify reports that River now participates in one out of every eight merged pull requests.Ramp’s Inspect extends the pattern further by executing inside sandboxed virtual machines that closely mirror developers’ local environments. The platform integrates with Datadog, Sentry, LaunchDarkly, Buildkite, and other engineering systems while remaining model-agnostic. Inspect supports multiple frontier models alongside MCP servers, custom tools, and proprietary workflows.Despite differences in implementation, all three platforms converge on the same architecture: durable execution environments, secure access to enterprise systems, workflow orchestration, and automated verification before code reaches production.Enterprises are insourcing the harnessThis is not unlike what happened with internal developer platforms over the past decade.Few enterprises built their own cloud infrastructure after adopting AWS or Azure. Instead, they built opinionated platforms on top of public cloud services that standardized deployment pipelines, security policies, and operational practices.The same pattern is now emerging around AI.The language model becomes another infrastructure dependency, while the enterprise-owned harness determines how that intelligence is applied. Routing requests, enforcing permissions, injecting organizational context, executing tools, and validating results become responsibilities of the platform rather than the model provider.Model selection consequently becomes a configuration choice instead of a workflow decision made by individual developers.Cost optimization becomes a platform capabilityCoinbase’s experience illustrates why enterprises increasingly want to control this layer.According to Chintan Turakhia, Senior Director of Production Engineering for Wallet and AI Engineering at Coinbase, the company’s infrastructure now processes more tokens than before while reducing overall spending.Coinbase co-founder and CEO Brian Armstrong attributes the improvement to multiple architectural optimizations rather than a single breakthrough. Coinbase introduced intelligent routing, lower-cost default models, prompt caching, smaller context windows, and better visibility into AI spending. The company has also experimented with open-weight models such as GLM 5.2 and Kimi K2.7.These optimizations are only possible because Coinbase owns the gateway sitting between developers and foundation models.Instead of asking thousands of engineers to switch tools, platform teams can update routing policies centrally. Model upgrades, pricing changes, and experimentation become infrastructure operations rather than organization-wide migrations.Commercial coding assistants still have a roleOwning the harness has not eliminated demand for commercial coding assistants.Claude Code remains the most widely used AI coding assistant among Coinbase engineers alongside Forge. Shopify developers continue using Claude Code and Codex in addition to River. Anthropic also lists Ramp among its enterprise customers, where nearly half of the engineering team uses Claude Code weekly.The emerging division of responsibilities is becoming clearer.Internal agents primarily handle asynchronous engineering workflows that begin in Slack or issue trackers, interact with internal systems, and validate changes against production telemetry.Commercial assistants continue to dominate interactive development sessions in which developers work directly within their editors or terminals.Rather than replacing each other, the two approaches complement different engineering workflows.Platform ownership matters more than model ownershipFor most enterprises, the question is no longer whether to build or buy an AI coding assistant.The more important decision is whether to own the orchestration layer that sits between developers and foundation models.Organizations with mature platform engineering teams can justify investing in proprietary harnesses because they gain centralized governance, model portability, security enforcement and cost optimization. Smaller engineering organizations are likely to continue relying primarily on commercial tools, where vendors absorb the operational complexity.RequirementRecommended optionRationaleAsynchronous work that must query internal databases, feature flags, and telemetry to verify itselfInternal agentProprietary integrations, credentials, and verification steps become first-class rather than extensions bolted onto a vendor clientInteractive work on a familiar repository at the editor or terminalCommercial assistant such as Claude Code or CodexEngineers already know the ergonomics, and the vendor absorbs the operational complexityChanging the default model across a large engineering organizationInternal gateway in front of bothRouting and spend policy move once, without asking engineers to change toolsTeams without a dedicated platform engineering groupCommercial assistant with per-developer quotasThe investment behind Forge, River, and Inspect assumes people whose job is to maintain itThis mirrors the evolution of cloud platforms, where enterprises selectively built higher-level abstractions without attempting to replace the underlying infrastructure providers.AI economics remain unpredictableOwning the harness does not eliminate infrastructure costs.Walmart recently introduced usage limits for its internal coding assistant after demand exceeded available budgets. Earlier this year, Uber exhausted its annual AI coding budget within four months, while GitHub Copilot has shifted to a usage-based credit and developer-quota model.Recent research from Stanford University and Microsoft Research helps explain why these costs remain difficult to predict. Their analysis of agentic software engineering tasks found that autonomous coding workflows can consume up to one thousand times more tokens than interactive code generation. Even identical tasks frequently exhibited order-of-magnitude differences in token consumption, making budgeting significantly harder than in conventional cloud infrastructure.Cost optimization therefore becomes another responsibility of the platform layer rather than the model itself.The platform becomes the strategic assetThe most significant architectural shift is that enterprises are beginning to treat AI agents as part of their internal developer platform.A decade ago, organizations differentiated themselves through deployment pipelines, infrastructure automation, and platform engineering. Today, they are extending the same philosophy to AI-assisted software development.The language model increasingly resembles interchangeable infrastructure. The enterprise-owned harness becomes the control plane that governs context, security, execution, verification, and economics.Model providers will compete to become the preferred reasoning engine inside platforms they do not control.If common agent runtimes and open protocols continue to mature while enterprises keep their policy and context layers proprietary, competition will increasingly move above the foundation model. Model providers will compete to become the preferred reasoning engine inside platforms they do not control.That may prove to be the defining architectural shift of enterprise AI. The strategic advantage lies in moving away from the model itself and toward the platform that determines how, when, and where it is used.The post Coinbase, Shopify and Ramp all built their own coding agents. All three still pay Anthropic. appeared first on The New Stack.