AI coding agents are now integral members of the team across most of the top software companies.
At Stripe, agents merge more than 1,300 pull requests a week. The company says those pull requests contain no code written by a person, and people review every change.
The top coding models today are built by OpenAI, Anthropic, and Meta, and while model companies provide their own harnesses, many software teams are shifting towards model-agnostic platforms like Codespeed. These platforms empower a team to swap to the best model as soon as an upgrade is available while keeping work in the same place.
A core component of great agents is memory, and it's still in the very nascent stages of innovation. While most agents today run rudimentary systems with durable memories added to tracking files, the industry as a whole is quickly shifting towards much more intelligent forms of… intelligence.
This intelligence is made up of the vast signals, feedback, and many more infinite touch points that train your agents' intelligence, and it's extremely valuable. But if you're continuously switching harnesses, or working on a platform that doesn't support multiplayer, then that memory and intelligence is unstable and inconsistent.
Instead, a new ideal is emerging among the best software factories. Pioneering forms of intelligence like Codespeed Intelligence focus on a model-agnostic system built around your team's preferences, taste, and judgment.
Microsoft believes agent security and governance should live outside the models
What makes an agent specialized to your company is everything your team has built up around it, but there's long been a debate about where training data and this type of intelligence should live, and whether it should be shared with model companies.
On October 10, Microsoft CEO Satya Nadella argued from a security perspective. In a post seen more than twelve million times, he argued that companies should treat frontier models as insider risks, just like operational security policies for employees.1
Nadella’s concern is that nobody can see inside a model. With traditional software, engineers could trace any behavior back to the code that produced it. With a model, nobody can trace a behavior back to the data it learned from or the way it was built, meaning they can't reliably know exactly why an agent behaved the way it did, and yet companies are already handing models their data and letting them act on their own. His answer is to keep the controls outside the model: what it can access, what it is allowed to do, a log of everything it did, and a way to stop it mid-task.
We need to separate the supply of intelligence from the authority over it.
And crucially, that separation implies a place outside a tool or harness delivered by the big model companies as well. He also wrote that no important outcome should depend on a single model. Whatever a company wants to keep needs to live somewhere that stays put when the model changes.
Teams will rent the model and keep everything else
For any team running AI coding agents, that leads to a simple rule: rent the model and keep everything else. Everything else includes the controls Nadella lists, the team's coding standards and architecture decisions, and the history of every ticket, pull request, and review. All of it should live somewhere the team owns, and any model should be able to work from it.
The tools the model companies sell break that rule. Each one works best with its maker's models, keeps its memory and history inside its own product, and ties that memory to one person instead of the whole team. When another lab ships a better model, a team using one of those tools has to choose between staying put and starting over.
That's the single-model dependency Nadella warns against.
Companies and their teams will end up treating models the way they treat cloud providers today. They'll pick the best one for each job and switch when a better one fits the job, keeping what makes their setup valuable and unique completely outside of the models and their harnesses.
Codespeed is built for this. On Codespeed, people and agents work together in one shared workspace with the controls Nadella describes, and everything stays with the team regardless of the model everyone is racing towards in a given week.
