Per-developer environments were the goal. Agents moved the goalposts.
The New Stack Arjun Iyer
Agents broke the old rule that one developer = one workstream. Now the thing that needs isolating is each change, not each person.
Based on reporting by The New Stack, Arjun Iyer — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
For years, platform engineering kept shrinking the tenant. Mainframes shared one machine across departments. Virtual machines gave teams their own slice. Containers and Kubernetes namespaces got it down to the developer. That was the target in the 2020s: one isolated environment per person, sized by headcount and planned on the assumption that one human was driving one stream of work at a time.
Coding agents blew a hole in that assumption. A single developer can now have several agent sessions running at once, and each of those sessions can be pushing a different change through validation. Anthropic’s engineers reportedly ran nearly 2,000 Claude Code sessions over two weeks while building a C compiler with parallel agents. Cursor’s docs even encourage running as many agents in parallel as you want. The worker is no longer the unit that matters.
The article’s point is blunt: the tenant has shrunk again, and the new tenant is the change. That matters because the pressure on the platform now scales with changes in flight, not with employee count. A Microsoft study found developers merged about 24% more pull requests over four months after adopting command-line coding agents, but merged PRs are only the visible part. Every merge has a trail of iterations, abandoned attempts, test runs, and temporary state behind it.
That changes the math fast. A 50-developer team supervising a few agent sessions each can create hundreds of active changes on a busy day. Each one needs its own writable data, its own view of shared topics, and its own running version of the services it touches. Seat-based planning starts to look absurd when one seat is really covering five workstreams.
The fix, according to the piece, is to treat each change like a tenant in a proper multi-tenant system: cheap to create, isolated only where it changed, and torn down when the change merges or gets abandoned. The source points to copy-on-write database branching from Neon and Xata as one way to make that cheap enough. The broader message is pretty simple: if platform teams keep sizing environments for people, they’ll keep building queues for software.
My take — AI-written commentary, not fact-checked reporting
The industry keeps trying to shove agents into human-shaped infrastructure, and that’s exactly backwards. Person-sized tenancy was already a convenient fiction; agent-native SDLC just made the fiction expensive. Anyone still budgeting by seats is basically bringing a paper umbrella to a cloud burst.
Read more about this at: The New Stack