Coinbase, Shopify and Ramp all built their own coding agents. All three still pay Anthropic.

The three companies have shifted from using off-the-shelf AI tools to building bespoke coding agents that sit on top of Anthropic's language model. This gives them control over prompt engineering, integration with internal CI/CD pipelines, and the ability to enforce company-specific policies. However, the agents still depend on Anthropic's API for the actual code generation, so the core capability is not owned in-house.

Adopting this hybrid approach incurs two cost streams: the ongoing fees paid to Anthropic for model usage and the internal expense of designing, implementing, and maintaining the custom agents. Engineering resources must be allocated to keep the orchestration logic up to date with evolving development workflows. Budgeting therefore needs to account for both external licensing and internal development overhead.

The agents' functionality is bounded by Anthropic's service availability and any usage limits imposed by the provider. If Anthropic experiences downtime, changes its pricing, or updates its API, the internal agents could fail or require rapid rework. Consequently, reliability planning must include fallback strategies for the LLM dependency.

Seeing multiple large firms arrive at the same architecture suggests a de facto standard for AI-assisted software development in enterprises. Teams at other organizations may view this pattern as a template for balancing customization with the proven performance of a commercial LLM. The trend also signals that vendor lock-in remains a consideration even when substantial internal tooling is added.

Overall, the shift to custom coding agents layered on Anthropic's model changes the engineering landscape: it adds a layer of internal complexity, introduces new cost categories, and retains a critical external dependency. Engineers must weigh the benefits of tailored workflows against the added maintenance burden and risk of service interruptions.