|
February 13, 2026
Fleets of agents humming across a software assembly line don't fit on a laptop. At Stripe, Ramp, and Spotify, agents now ship thousands of PRs per week. What's blocking everyone else is what has been broken all along: your development environment.
I wrote about why background agents are ending localhost and what it means for your team: The last year of localhost →
|
Featured
|
This is not a future prediction, but reality today. Last week, Ona authored 88.5% of the PRs we merged on main. Scheduled agents pick up tickets at 7am and have PRs ready by standup. Background agents triggered from Sentry triage and fix bugs without anyone having to initiate the process. Because our environments run inside our own cloud account, agents can query a replica of our production database, check Grafana dashboards, trace errors in Honeycomb, and hit internal APIs to debug and validate their own work.
Stripe
Stripe built Minions, their one-shot coding agents, which now merge over a thousand agent-authored pull requests per week. What's notable is that the breakthrough wasn't a better model. It was that every Stripe engineer already had a cloud-based devbox with full monorepo checkout, test suites, and credentials. The agent infrastructure turned out to be a thin layer on top of years of environment investment.
Ramp
Ramp arrived at the same place through a similar path. Their background agent now accounts for 57% of all merged PRs, and it was built on standardized environments that existed before the agent did. Environment standardization first, agents on top. The infrastructure investment predated the agent era by years, and it's now paying compound returns.
Spotify
Spotify's Co-CEO Gustav Söderström said on their Q4 earnings call that the company's best developers haven't written a single line of code since December. Their internal system "Honk" lets an engineer fix a bug from Slack on their phone during their morning commute and merge it to production before they arrive at the office. They shipped over 50 new features in 2025.
Cloud development environments are the prerequisite
None of these companies built a special agent harness or found a smarter model. What they had was standardized, reproducible, cloud-based development environments. The gap between "generates a diff" and "opens a merge-ready PR" is, and has always been, the development environment. If your agents can't run the application, execute tests against real services, or validate their own work, you're unlikely to get the productivity lift you're hoping for, regardless of how capable the model is.
With the arrival of background agents, the move to the cloud is finally happening. Through a different door than we expected.
We wrote about this in detail. If you're thinking about where to invest next, we think it's worth reading:
What the post covers
Why git worktrees break.
If you're a developer productivity engineer tasked with running three agents in parallel, you've probably started with git worktrees. In a monorepo, this breaks immediately. Each worktree needs its own dependency install, its own running services, its own database instance. Port conflicts, shared caches corrupting each other, and a machine that grinds to a halt.
Isolation: VMs, not containers.
Agents run arbitrary code remotely. Containers share a kernel with the host. A container escape gives an attacker access to every other container on the same machine. The right primitive is a virtual machine, where each environment gets its own kernel, memory space, and network stack.
Declarative environments.
The Dev Container spec is the underappreciated hero of this story. A devcontainer.json codifies everything an environment needs. Given this file, any machine produces an identical environment. We created .gitpod.yml years ago, learned from that, and adopted Dev Containers as the foundation when we rewrote our architecture from scratch.
Connectivity and context.
An agent's output is proportional to the context quality it can access. An agent running inside your network can query your databases, hit your internal APIs, and run your full test suite against staging. Most agent sandboxes punt on this.
Security: assume compromise.
Anyone who tells you they've "solved" prompt injection is selling something that doesn't exist. The right question isn't "how do we prevent compromise?" but "what can the agent reach when it is compromised?" We monitor every system call, file access, and network packet at the kernel level.
The compounding effect.
When a development environment is one click away, product managers, designers, and support teams can access the codebase directly. A designer adjusting spacing can spin up an environment, have an agent make the change, and send a PR instead of pulling an engineer out of deep work.
Try Ona
If any of this resonates, the best way to understand it is to try it. Give Ona two minutes and your most dreaded ticket, and see what comes back. Start now →
|
May your PRs merge without the dreaded 'small nit',
Johannes
CEO Ona
|