AI coding agents: who actually controls your software now?

Coder, the development-environment platform already used by banks and public administrations, announced Agent Relay on September 2, 2026: an integration with SpaceXAI that runs Cursor’s cloud coding agents inside the customer’s own infrastructure, so source code, credentials, and internal systems never leave their network (The Next Web; Coder’s press release). At Evicron, an AI and custom software studio based in Barcelona, we’ve spent months watching our own clients trial coding agents and run into the same question before signing off: if an agent writes, runs, and ships code on its own, who is actually in control of that software?

What Coder and SpaceXAI just shipped

Agent Relay solves a very specific problem. Tools like Cursor’s cloud agents are convenient because the developer keeps working from the same app or browser they already use, but until now the agent’s reasoning — and the tool calls it makes against your codebase — ran on the provider’s own infrastructure. For a bank, a government agency, or any company bound by data-residency rules, that was reason enough to keep these tools off-limits. With Agent Relay, Cursor still handles the model’s planning, but the actual execution against your code happens on machines the customer controls, inside their own network. It’s currently in early access with design partners, so it isn’t a general-availability product yet.

What matters here isn’t really the product itself — it’s the signal. A coding-agent vendor as established as this one is admitting that where the code and the secrets physically live is, for a large share of its prospective customers, the deciding factor on whether they adopt the tool at all.

Why this isn’t just a banking problem

It’s tempting to assume this only matters for heavily regulated industries. It doesn’t. Any SME that lets an AI agent write, modify, or deploy parts of its software is handing a technical decision to a system acting with real autonomy: it can touch a production database, pull in a new dependency, or leak a credential in a commit before anyone reviews it. The EU AI Act and the GDPR already require companies to know, at all times, where their data is processed and who can access it — and a coding agent working against your repository and your environment variables falls squarely under that obligation, whether it’s a bank or a ten-person retailer using it.

The practical question isn’t whether to use coding agents — avoiding them entirely is getting harder, and the productivity gains are real — it’s under what guarantees.

Questions your company should answer before approving a coding agent

1. Where does the code run, and where do your secrets live? If the answer is “on the vendor’s cloud” and your business handles sensitive data or contracts with residency clauses, you need to know that before signing, not after an audit.
2. Is there a record of what the agent actually did? Every commit, every new dependency, every deployment an agent triggers should be traceable the same way a human teammate’s work would be.
3. Does a human review changes before production? An agent proposing a change is fine; letting it approve and ship that change unsupervised is a risk very few companies should take on yet.
4. Who owns the resulting code? With low-code tools or closed platforms, migrating later can get expensive. The more standard the stack underneath the agent, the easier it is to switch vendors if you need to.

How we approach this at Evicron

In our custom software projects we use AI to speed up parts of development, but the code we hand over belongs to the client, with no vendor lock-in clauses, and every change goes through review before reaching production: weekly sprints, a clickable demo from week two, and a clear trail of what got built and why. When the question runs deeper — which agents your company should approve, what internal policy you need before letting one touch real code — that’s where our AI consulting for businesses comes in: we audit which tools your team is already using, sanctioned or not, and what controls are missing to make them safe.

Bottom line

A vendor as established as Coder shipping a version of its agents that runs inside the customer’s own network confirms something many companies already suspected: the convenience of a coding agent isn’t worth losing track of where your software lives and who can touch it. Before approving one at your company, work through the four questions above; if you don’t have an answer for one of them, better to have it before the agent writes its first line of code than after.

Want to know which AI tools your team is already using, and under what actual guarantees? Get in touch: the first discovery session is free and we reply within 24 hours.