AI labs move to own the entire software development workflow

The profit motive for integration

Running a high-end AI lab is an expensive endeavor. While serving existing models is profitable, the cost of training the next generation of frontier models is massive. As analyst Benedict Evans noted, Anthropic faces a challenge where over 80% of its third-quarter revenue came from models released in just the previous six months.

To improve margins, these companies need to extend the lifespan of their current products. If a lab can go 18 months between major training runs instead of six, the business becomes significantly more sustainable. This pressure is driving Anthropic and OpenAI to move beyond just offering smart models to owning the actual workflows developers use daily.

By capturing more of the development process, these labs hope to create stickier products. The goal is to move from being a tool that a developer occasionally calls to being the primary environment where the work happens. This strategy aims to turn a temporary preference for a specific model into a long-term commitment to a company's entire software suite.

Owning the developer's workspace

OpenAI recently announced an agreement to acquire Pierre, the company behind Code Storage. This move is intended to help OpenAI's Codex model interact more naturally with how developers browse files and review changes. The company wants Codex to be present from the very first branch an agent creates until the final deployment of an application.

Anthropic is pursuing a similar strategy with its redesigned Projects feature. This tool allows a coordinator to parcel out work to several parallel Claude Code sessions that share a single memory. This shared memory can store specific details like why a feature was cut or who needs to be consulted before touching a specific service.

These features allow AI agents to access the 'tribal knowledge' of a human team. A maintainer might remember a specific customer's needs for an old API, while a security engineer knows why a certain change requires a second review. By capturing this information in a project record, the next agent can find it and act on it without human intervention.

The risk of losing neutrality

Historically, the software industry has thrived on neutral spaces like GitHub, where different tools can work together. For example, OpenAI built a custom Git back end called Pando to handle concurrent reads and writes, but they kept it compatible with standard Git protocols. This allowed their engineers to use familiar tools while the underlying infrastructure was modernized.

Neutrality allows for competitive differentiation and easier switching. If a developer uses a neutral platform, they can send a patch from a Codex-powered agent to a maintainer who uses Claude. The two can discuss the change using the project's tools and run it through the same automated checks regardless of which AI they prefer.

When labs move to own the workspace, this flexibility disappears. If only one specific agent can access the most useful history of a project, the supplier of that agent gains a massive advantage. This makes it much harder for a team to try a rival startup's tool if that tool isn't allowed to see the project's authorized history or data.

Competing with integrated giants

The labs are entering a space where companies like Apple already have a foothold. Developers often accept Apple's strict rules because the integrated environment makes it easier to build and ship software for their customers. The frontier labs are trying to build their own versions of these integrated environments where the model and the tools are designed to work perfectly together.

However, even integrated giants often find value in opening up to rivals. For instance, GitHub recently announced a preview of Copilot inside Apple's built-in coding assistant. This allows Apple to keep its platform while still giving developers access to the best tools available on the market.

The future of the industry depends on whether these labs will create closed ecosystems or open ones. If a team likes a specific workspace but wants to use a different agent for a specialized task, they should be able to do so. The ideal scenario is one where labs provide the power, but neutral spaces allow developers to choose which specific tools they use for each job.

What to watch next

The next major milestone will be seeing how these labs handle multi-agent collaboration within their own walled gardens. We should watch for how Anthropic's Projects feature handles conflicting instructions from different Claude sessions.

Additionally, keep an eye on how GitHub responds to these moves. They may need to find new ways to ensure that third-party agents can remain competitive even as OpenAI and Anthropic try to build more comprehensive, end-to-end workflows.

Finally, watch for the first major