Relace wants to build the infrastructure foundation for AI coding agents after raising $23M in funding

Artificial intelligence coding infrastructure startup Relace today announced it has closed on $23 million in early-stage funding.

The Series A round was led by one of Silicon Valley’s most prestigious venture capital firms, Andreessen Horowitz, and saw participation from Matrix Partners and Y Combinator.

The startup, which is officially known as Squack Inc., is building specialized infrastructure for AI coding agents, which are already generating massive amounts of code and need a new set of tools and services to operate effectively.

In a blog post announcing the round, Relace co-founders Preston Zhou (pictured, left) and Eitan Borgnia (right) said they’re building the models and infrastructure required to enable “software on demand”. What that means is small and highly specialized models that run on optimized infrastructure, so AI can generate code faster and ensure it’s production ready straight away. In other words, they want to create the same kind of fine-tuned development environment for AI coding agents that human developers already enjoy.

Andreessen Horowitz outline the problem Relace is trying to solve in a post on X:

For decades, code was written by humans. Now, it’s generated by agents.

The bottleneck has moved: from writing code to running it.@relace_ai is building the infra that makes coding agents production-ready

We’re proud to lead their $23M Series A.@EBorgnia @pfactorialz https://t.co/WbfPOGW8JY

— a16z (@a16z) October 8, 2025

Chief Executive Officer Zhou has a strong background in AI research and engineering, while Chief Technology Officer Borgnia brings deep experience in systems programming and infrastructure. Together, they have formed a small, eight-person team that’s focused on building low-latency, cost-effective models that are purpose-built for AI coding.

In the blog post, the co-founders explained that there are dozens of AI startups all scrambling to try and build the best AI coding agents, and what they’re seeing is that many of them are, in fact, quite similar. “Each is independently rebuilding the same toolsets, sandboxed execution environments, and source control systems to power their products,” they wrote.

They believe that as the AI coding industry evolves, these kinds of agent design choices, which are currently considered as differentiators, will eventually become managed services, in the same way that databases, deployments and authentication are all outsourced today.

But it won’t be easy to build these services, they say. “This infrastructure requires unique tradeoffs and must be co-optimized alongside models for performance and speed,” they wrote.

To that end, Relace has forged the beginnings of an infrastructure foundation for AI coding agents that’s based on three core model types. These include “apply models” that integrate AI-generated code directly into live projects without human cleanup, and “embedding models” that enable coding agents to search for, and retrieve relevant code snippets from massive codebases in seconds. They also include “reranking models” designed to filter multiple AI outputs and select the most accurate one, in order to reduce the frequency of hallucinations, or inaccurate outputs. Zhou and Borgnia explained that these models can be plugged into general-purpose large language models or specialized code generating models to enhance their coding skills, so the code they generate performs better and has fewer bugs.

Beyond these models, Relace is also building out infrastructure for the various stages in the development lifecycle, including versioning, deployment and codebase state management. Zhou and Borgnia believe it’s needed, because traditional developer environments were built for humans. They say coding agents need frameworks that allow them to slide directly into production environments, so they can operate with greater autonomy.

Relace’s combination of specialized models and coding agent infrastructure has already been deployed by AI companies such as Lovable Inc., Magic Patterns and Orchids, dramatically improving the efficiency of their coding agents. “We can accurately surface the necessary codebase context for agents in ~1-2s with our specialized embedding & reranker models, and we merge file edits at over 10,000 tokens per second with our apply model,” the co-founders said.

The startup is now in expansion mode, and will use the funds from today’s round to grow its engineering team, create more specialized models and target more enterprise customers to help them experiment with agentic workflows.

Image: Relace