Connor Loi and Saai Arora might be the first founders to genuinely mean it when they say they "eat their own dog food."
The two engineers behind Replicas claim that artificial intelligence agents—not human developers—have authored roughly 95% of the code shipped since they started the company. Not as a proof of concept. Not for a demo. As the actual way they build their product, day in and day out. It's a claim made by the founders themselves, without independent verification.
It's the kind of statistic that sounds almost reckless until you consider what Replicas actually does: it markets the same workflow to other startups. And according to Loi, it's working. Teams that adopt the platform report that AI agents now generate about 30% of their pull requests—a self-reported statistic from the company—a figure that would have seemed absurd even two years ago.
The company came out of Y Combinator's Spring 2026 batch and launched its first full platform version in March 2026. The premise is deceptively straightforward. Developers want AI that can autonomously pick up tickets, write code, and open pull requests. But most engineering leaders won't let an AI agent anywhere near production repositories without serious guardrails in place.
Replicas built the scaffolding to make that delegation feel safe.
Agents You Already Pay For
Here's where things get interesting. The platform doesn't force you into a proprietary model. Instead, it runs on top of Claude Code or Codex—AI coding agents you connect using your own subscriptions or API keys.
Each task the agent picks up spins up an isolated virtual machine, clones your repository, installs whatever dependencies your project needs, runs databases and services, completes the work, then opens a pull request. The sandboxed environments auto-delete after a week of inactivity. Your code never leaves the connected repository. The company doesn't train on customer data. Only you have SSH access to the workspace.
In June 2026, according to their changelog, Replicas went further and restricted public network access to the sandboxes entirely. The security documentation gets updated frequently—sometimes more than once a month.
Whether this architecture actually solves the trust problem or simply moves it around remains an open question. But the approach has found early believers.
Bots as Teammates
Using Replicas feels less like prompting an AI and more like delegating to a junior developer you never have to onboard.
Mention @Replicas in a Slack channel. Assign a Linear issue to the bot the way you'd assign it to a teammate. Drop a comment on a GitHub issue tagging @tryreplicas. The agent picks up the task and gets to work.
The feedback loops run without supervision, at least in theory. When CI checks fail on an agent-authored pull request, Replicas reads the logs and iterates. Same with code review comments. Organizations can configure policies to block agents from merging their own PRs, require drafts by default, or limit which repositories are accessible.
For teams juggling multiple projects or wrestling with monorepos, the company introduced "Environments"—preset configurations that bundle environment variables, files, startup hooks, and something called Model Context Protocol skills. Warm pools pre-initialize workspaces to cut latency. Preview URLs let you inspect services running inside the sandbox, like a web server on a given port, directly from your browser without SSHing in.
The experience starts to feel a bit like managing cloud infrastructure. Which, in a way, it is.
The Automation Creep
By mid-year, Replicas had added cron triggers, Sentry integrations, and webhook-based automations with per-automation overrides. You can specify which agent, which model, even which workspace size a particular automation should use. Workspace configurations range from small (2 vCPU, 8 GB RAM, 20 GB disk) to large (4 vCPU, 16 GB RAM, 32 GB disk).
The company offers a REST API with org-level and personal keys, endpoints to create replicas and send messages, server-sent events for streaming responses, and webhooks. But programmatic use is sanctioned only through the API or official automation flows. The terms of service explicitly prohibit scripting the dashboard or integrations—a clause that suggests someone, somewhere, tried.
Racing a Crowded Field
When Loi and Arora launched on Y Combinator's community forum in March, they mentioned that more than 20 YC startups were already using the product at that time. Named customers include Mintlify, Knowunity, and Composio. Whether that early cohort has grown significantly since isn't clear, though the 30% pull request figure Loi cited in June suggests at least some teams are leaning in.
The market they entered is both crowded and still figuring out what it wants to be. GitHub Copilot has edged toward agentic workflows. Cursor emphasizes editor-native agents that live where developers already work. Replit pushes "agent-first" cloud development environments. Coder announced self-hosted agents earlier this year for teams that won't send code to the cloud under any circumstances.
And then there's Stripe. In February, the payments giant published details about "Minions," its internal unattended coding agents running in isolated cloud environments. The architecture mirrors what Replicas offers, but built in-house and not for sale. That kind of in-house development from a major tech company can validate a market or make it harder for startups to gain traction. Sometimes both.
Replicas differentiates on two fronts: the bring-your-own-agent model and the emphasis on background execution over interactive sessions. You delegate work to the platform, then come back later to review pull requests. No sitting in a chat thread watching the agent "think" through its reasoning steps in real time.
Whether that's a feature or a limitation depends on how much trust you're willing to extend.
Still Moving Fast
The two-person team in San Francisco ships updates almost daily. Recent changelog entries include fixes to Claude session reliability, improvements to the "thinking" state UI, and account-level controls for how pull requests get attributed.
GitLab OAuth support is rolling out gradually, currently behind a feature flag. OpenCode support is listed as "coming soon" on the homepage, which would add a third agent option. The pricing page references SOC 2 compliance for the Enterprise tier, though no public audit report has been posted.
For now, Replicas remains a small, fast-moving operation betting that the future of software development involves far more delegation and far less hands-on coding.
If their own 95% statistic holds—and it's worth noting that claim comes from the founders themselves, not an independent audit—they're already living in that future. The rest of the industry is still deciding whether to follow.
