In the ever-evolving landscape of AI-assisted software development, a fascinating trend is emerging. Major players like Coinbase, Shopify, and Ramp are not just adopting AI tools; they're building their own coding agents. But here's the twist: they're still paying Anthropic. This raises a deeper question about the future of AI in enterprise settings.
The Rise of Internal Coding Agents
These companies have taken a unique approach, developing internal coding agents to assist their developers. However, it's not a simple case of build-or-buy; it's about architectural choices. The key layer they've chosen to own is not the language model itself but the 'agent harness' - a critical component that provides context, permissions, and workflow management.
The Power of the Harness
The agent harness is like the conductor of an orchestra, directing and coordinating the various elements. It's this harness that determines how the large language models from companies like Anthropic, OpenAI, and Google are utilized. By owning this layer, enterprises gain control over the entire development process, from context and permissions to tool access and verification.
Open-Sourcing the Architecture
LangChain's recent move to open-source Open SWE is a significant development. It provides a public implementation of the architecture pattern already in use at Stripe, Ramp, and Coinbase. This move suggests that this architecture is not just a one-off solution but a repeatable, scalable model.
Competitive Advantage
The competitive edge in this scenario lies beyond the language model. It's in the unique ways enterprises utilize and integrate these models into their development processes. Take Coinbase's Forge, Shopify's River, and Ramp's Inspect - all independently developed yet remarkably similar in their problem-solving approaches. They integrate with various tools and systems, providing a seamless workflow that enhances developer productivity.
Insourcing the Harness
This trend is reminiscent of the shift towards internal developer platforms over the last decade. Just as enterprises built opinionated platforms on top of public cloud services like AWS and Azure, they're now doing the same with AI. The language model becomes a dependency, while the enterprise-owned harness dictates its application.
Cost Optimization and Control
Coinbase's experience is a case in point. By owning the gateway between developers and foundation models, they've achieved significant cost savings and improved efficiency. This level of control allows for centralized management of routing policies, model upgrades, and pricing changes, without disrupting individual developers.
The Role of Commercial Assistants
Despite the rise of internal coding agents, commercial assistants like Claude Code and Codex still have a place. They dominate interactive development sessions, where developers work directly within their editors or terminals. The two approaches complement each other, each serving different engineering workflows.
Platform Ownership
For enterprises, the question is not whether to build or buy an AI coding assistant but whether to own the orchestration layer. Mature platform engineering teams can justify investing in proprietary harnesses, gaining centralized control and optimization. Smaller teams may opt for commercial tools, where vendors manage the complexity.
AI Economics
While owning the harness provides control, it doesn't eliminate infrastructure costs. Recent examples like Walmart and Uber highlight the challenges of predicting and managing AI costs. Research from Stanford University and Microsoft Research explains why these costs are so unpredictable, with token consumption varying significantly across tasks.
The Strategic Shift
The most significant shift is the integration of AI agents into internal developer platforms. Enterprises are treating AI as an extension of their development infrastructure, with the language model as interchangeable infrastructure. The real strategic asset is the enterprise-owned harness, which governs the entire AI-assisted development process. Model providers will compete to be the preferred reasoning engine within these platforms.
Conclusion
As AI continues to evolve, the architectural decisions made by enterprises will shape the future of AI in software development. The strategic advantage lies in owning the platform that determines the use of AI, not just the AI models themselves. This shift could define the landscape of enterprise AI, with model providers vying for a place within these proprietary platforms.
What makes this particularly fascinating is the potential for a new era of AI competition, focused not on the models themselves but on their integration and application within enterprise systems.
