The next AI race is happening around the model

Companies are building the environment around AI

Shopify has an AI agent in Slack that reads code, runs tests and opens pull requests. Underneath it sits a platform designed to let the company keep changing how its agents work. One of the agents it supports runs on Pi, an open-source harness now housed at Earendil. Shopify engineering, Pi’s Earendil announcement.

That is a more interesting development than another argument about which model writes the best code.

Companies are starting to build the environment around AI: the place it gets its instructions, the systems it can reach, the work it remembers and the checks it has to pass. The model still matters enormously. But more of what makes an agent useful belongs to the company running it.

The temptation is to describe this as companies moving away from Claude. That would miss what is actually happening. Coinbase, for example, says its engineers use internal tools alongside Claude Code, Cursor and OpenCode. Building your own agent system can still mean buying intelligence from the same providers. Coinbase engineering.

What an agent harness lets you change

A model and an agent harness do different jobs. The model generates responses and proposes actions. The harness supplies context, calls tools and manages the ongoing exchange between the model and its environment. A larger platform can add persistent records, access controls, isolated execution and connections to the rest of the business.

Pi makes that distinction tangible. It is a small, extensible agent harness with support for multiple model providers. Developers can add tools, change how context is handled and embed it in their own applications. You can use Anthropic models through it. You can also use other providers without adopting an entirely different agent interface. Pi documentation.

Open source gives teams somewhere to start and code they can change. It does not make every engineering decision for them. Pi’s own repository says it has no built-in permission system restricting filesystem, process, network or credential access. Those boundaries need to come from the environment and controls around it. Flexibility brings work with it. Pi repository.

Shopify and Coinbase are solving different problems

Shopify’s implementation gives a sense of where that work can lead. Its platform, Aquifer, separates the durable conversation from the agent runtime and the sandbox where code executes. The conversation survives when a process disappears. The runtime can be replaced. River, the agent people talk to in Slack, is one configuration on that platform; a headless Pi agent is another. Under the River.

The interesting possibility is reuse. Once a company has a dependable place for agents to work, each new use case may require less plumbing. A code-review agent and a research agent need different instructions and access. They should not necessarily need two completely separate systems for storing their work and tracking what happened.

Coinbase arrived at a related problem from another direction. Its engineers had coding agents, but much of the surrounding workflow remained sequential. Mux gave agents separate workspaces so engineers could coordinate tasks in parallel. Coinbase says the value of building it internally came from integrating its own deployment systems, review flows and repository conventions. It also says adapting existing open-source tools to its security requirements would have required substantial work. How Coinbase built Mux.

That complicates the idea that one open-source project is about to become everybody’s answer. Shopify and Coinbase are making different choices. Both are investing in software that reflects how their organizations actually operate.

Legal expertise becomes part of the system

The legal industry is taking this further. Thomson Reuters, the legal content and technology company behind Westlaw, has developed its own model, Thomson. It started with open-weight foundations and applied further training using proprietary professional content and expert-created data. That is a different investment from building an agent harness: it changes the model itself. The company says CoCounsel remains a multi-model product, using Thomson where it has an advantage and other leading models elsewhere. Thomson technical overview, launch announcement.

There is a law-firm example, too. Kirkland & Ellis announced a proprietary platform with Palantir in June 2026, built around its knowledge and workflows for private equity fundraising. The firm describes connecting transaction history, obligations and legal expertise into an operational system. Its announcement supports the case for custom infrastructure; it does not establish that Kirkland trained its own foundation model. Kirkland & Ellis announcement.

These projects put years of accumulated knowledge to work in different ways. Some organizations are changing the model. Others are building the systems around it. I expect both approaches to spread as companies get clearer about what their existing AI tools cannot do.

The next wave reaches beyond engineering

Consider an agent preparing a customer quote. The language model might understand the request perfectly and still be unable to finish the job. It needs current inventory, the right pricing rules and a way to recognize an exception. Someone has to decide whether it can promise a delivery date or approve a discount. The next model release will not automatically resolve a disagreement between the warehouse system and the sales team.

Those are business decisions expressed in software. As agents get more capable, that software becomes a larger part of what a company can do with them.

There is room here for new products, too. Some businesses will build their own agent infrastructure. Others will buy a managed platform and customize the parts that matter. Providers will compete over how well they preserve context, connect applications and let customers control what agents can do. My expectation is that reliable operation will become a stronger selling point as more agents move from demonstrations into everyday work.

What should survive the next model release

Changing models will still take testing. A replacement can interpret instructions differently, choose different tools or miss an exception its predecessor handled. A common interface makes the switch easier to engineer; it cannot establish that the new behavior is acceptable.

For a growing business, the opportunity is to make its knowledge and workflows more durable than any single model choice. Keep the rules accessible. Keep records of what agents did. Build ways to test a new model against work the company understands.

When a better model arrives, the business should be able to put it to work with the customer history, permissions and operating knowledge it has already built. Nobody should have to teach the company how to be itself all over again.

Our agentic AI work connects models to the systems and rules a business runs on. Talk with us about a workflow you want to build.