Someone assigns a ticket in Linear. Hours pass. Then a pull request materializes—code written, tests run, CI failures debugged, all without a single engineer touching a keyboard.
It sounds like the setup for a cautionary tale about automation. Instead, it's Tuesday morning at a growing number of Y Combinator startups, where engineering teams have started treating an AI agent called Replicas much like they'd treat a junior developer: mention @Replicas in Slack, point it at a GitHub issue, then move on to something else. By the time you check back, the work is done.
More than 20 YC companies are already using the platform, according to Replicas itself. The claim that feels most provocative, though, comes from the company's own marketing materials: teams using Replicas now ship over 30% of their pull requests through the platform. That's not autocomplete. That's a fundamental rebalancing of who—or what—writes production code.
Whether that figure holds up across all customers remains unclear. Replicas hasn't disclosed customer-level data, and independent verification is hard to come by. But the narrative fits a pattern. At Ramp, an internal tool called Inspect handles roughly 30% of merged PRs. Stripe's "Minions" agents reportedly produce over 1,000 pull requests weekly. The numbers may vary, but the direction of travel doesn't.
How It Actually Works
Replicas emerged from Y Combinator's Spring 2026 batch—a small company with an ambitious pitch. The product positions itself as a "background coding agent," and the mechanics are relatively straightforward, if somewhat unsettling in their autonomy.
You start a task the same way you'd loop in a colleague: a Slack mention, a Linear assignment, a GitHub tag. The agent spins up its own sandboxed virtual machine, clones your repository, installs dependencies, writes the code, runs the test suite. If continuous integration fails, it debugs. If a reviewer requests changes, it revises. All of this happens in isolated environments that teams can configure with custom variables, startup hooks, and Model Context Protocol servers for extended tooling.
When the work is finished, a pull request appears.
The platform integrates directly into existing workflows—Slack, Linear, GitHub, plus a dashboard, CLI, and REST API for teams that want programmatic control. Preview URLs let you test backend changes running inside the sandbox before anything touches production. The system even supports "warm pools," pre-configured environments that eliminate cold starts for commonly used setups.
It's a long way from the early auto-complete tools that merely suggested the next line of code.
The Engineering Behind the Speed
Replicas has moved quickly, perhaps more quickly than you'd expect from a team of two founders. Connor Loi and Saai Arora—both University of Waterloo computer science graduates with previous stints at Cohere, Ramp, and Shopify—launched an initial version in December 2025. By January, they'd migrated the sandbox backend to Daytona, cutting startup time from around a minute to five seconds. MCP support landed the same month, letting Claude Desktop and other MCP clients spawn replicas programmatically.
March brought warm pools and preview URLs. A V1 whitepaper followed, formalizing the architecture. Google Workspace integration is rolling out now, feature-flagged, allowing agents to create and edit Docs, Sheets, and Forms.
The pace suggests either a highly focused team or more resources than the public record reveals. Y Combinator job listings indicate a team of two. LinkedIn suggests the company may have grown to as many as 10 employees. Funding details remain murky: multiple sources including CB Insights, Preqin, and Dealroom list conflicting data ranging from $125,000 to $500,000, with investors including Y Combinator and Pioneer Fund. The company hasn't disclosed its valuation or confirmed total capital raised.
The Price of Automation
The pricing model reflects just how compute-intensive this kind of background work can be. A free Hobby tier offers one-time allotments of 1,200 minutes for both human-initiated and automated tasks—enough to experiment, not enough to run a team.
The Developer plan costs $120 per seat monthly and includes 5,000 automation minutes. Jump to the Team tier at $300 per seat per month, and you get 15,000 automation minutes plus beefier sandbox resources: 4 vCPUs, 16 GB of memory, 32 GB of disk space. For companies shipping hundreds of pull requests a month, those minutes add up fast.
Still, the economics might pencil out. If an agent handles 30% of your PRs, what's the opportunity cost of having senior engineers write boilerplate CRUD endpoints or one-off bug fixes?
Not an Isolated Bet
Replicas isn't alone in this space, and that's perhaps the most telling part. The last six months have produced a cluster of similar tools, both from startups and established players betting their platforms on the same shift.
GitHub shipped Copilot Coding Agent and Agentic Workflows in early 2026, weaving agents across issues, pull requests, and VS Code. Cursor launched long-running agents in February. Cognition's Devin continues iterating on its pay-as-you-go model. The convergence suggests industry conviction, or at least a shared hunch that this is where the puck is headed.
What distinguishes this wave from earlier code-generation tools is the emphasis on unattended, full-context execution. These aren't assistants that wait for you to accept a suggestion. They're systems designed to take a vague task description, work autonomously in isolated environments, and deliver production-ready code. The difference matters.
So do the guardrails. Sandboxed VMs, CI integration, code review hooks, restricted API scopes—all are designed to contain the blast radius when things inevitably go wrong. And they do go wrong. A January 2026 analysis of 33,000 agent-authored pull requests documented common failure modes. December research highlighted reproducibility gaps in AI-generated code. Security researchers have flagged supply-chain risks in AI dev toolchains, which helps explain the industry's obsession with strict sandboxing and limited integration permissions.
The tools are getting better. They're also getting riskier at scale.
What Happens When Agents Ship a Third of Your Code
The roadmap Replicas hinted at in its April 2026 Launch YC post includes the features you'd expect: expanded automation, recursive agent instances (agents spawning sub-agents), further optimizations to warm pools and snapshotting. None of it is particularly surprising.
What's more interesting is the behavioral shift the platform represents. When engineering teams start shipping 30% of their PRs through an autonomous agent—if that claim holds—it suggests something beyond incremental productivity gains. It points toward a reorganization of how software gets built. Less time on routine fixes and feature implementation. More time on architecture, code review, high-leverage decisions.
Maybe that's a win. Or maybe it introduces new forms of technical debt—code that works but no human fully understands, written by a system that optimizes for passing tests rather than long-term maintainability.
The question isn't whether agents will write code. They already do, and they're getting better at it every quarter. The question is how quickly teams will hand over background tasks to systems that work independently, out of sight, while humans focus elsewhere. And what happens when the percentage climbs past 30%, past 50%.
That future, for better or worse, already has one foot in the door.
