LiveKit Bought Loophole Labs to Kill the Agent Cold Start

LiveKit, the San Francisco company whose infrastructure sits underneath a large share of production voice agents, has acquired Loophole Labs, a small virtualization startup built around its own hypervisor. The deal, announced September 24, brings Loophole's entire eight-person engineering team into LiveKit, with founder and CEO Shivansh Vij continuing to lead it. Financial terms were not disclosed.

The stated goal is specific: get agent startup times under three seconds. LiveKit says it plans to move Loophole's technology into production over the next several weeks, according to the company's announcement.

What LiveKit actually bought

Loophole Labs, founded in 2021 and backed by Amplify Partners, Cervin Ventures and Peak XV according to its website, spent five years on two unglamorous problems: how fast an application can start, and whether a running workload can move between machines without falling over. Its answer was to build its own virtualization stack, including a hypervisor called Substrate, per the acquisition release.

That stack powered two products. Architect is a Kubernetes availability platform that live-migrates running workloads across nodes, zones and clouds before spot instances are reclaimed, and says idle pods can hibernate and wake in under 50 milliseconds. AGX is an agent sandbox platform still in early access. Its site advertises a 2 GiB microVM booting in 0.74 seconds, sub-agent forks in 7 milliseconds, and snapshot restores in single-digit to low double-digit milliseconds, with a claim of booting and forking up to 10 times faster than Firecracker, the microVM technology behind AWS Lambda.

Vij said in the release that the two teams had worked closely together over the past year before making it official. That suggests LiveKit is not buying a roadmap; it is buying technology it has already tested against its own workloads.

Why cold starts are the whole ballgame for voice

A text agent that takes a few seconds to wake up is an annoyance. A voice agent that takes a few seconds to answer a phone call is a hang-up. LiveKit's customers, which the company lists as including Salesforce, Meta, Microsoft, SpaceXAI, Spotify and 911 emergency services, run conversations where latency is the product.

The traditional fix is to keep agents warm: pre-provisioned workers sitting idle, waiting for a call. That works, and it is expensive, because you pay for capacity that is mostly doing nothing. It also sets a ceiling on concurrency, since the warm pool has to be sized for peak demand. LiveKit's announcement names all three levers the acquisition is meant to move: responding faster to demand swings, cutting costs because agents no longer sit idle, and getting closer to what it calls essentially unlimited concurrent agents.

CEO Russ d'Sa put the requirement plainly in the release: agents need to start fast and adapt as usage changes. The Loophole approach, snapshotting a ready-to-run environment and restoring or forking it on demand, is the standard way to get there. Owning the hypervisor rather than renting someone else's abstraction is what lets LiveKit tune it for its own traffic.

The platform play behind the plumbing

The deal fits a pattern LiveKit has been running all year. When it raised a $100 million Series C at a $1 billion valuation in January, led by Index Ventures with Salesforce Ventures, Altimeter, Redpoint and Hanabi Capital participating, the company described itself as building the development stack and runtime between foundation models and end-user apps. Its lineup already included the open-source LiveKit Agents framework, which it said was downloaded more than a million times a month, plus Serverless Agents hosting, a no-code Agent Builder, an inference router, and observability tooling.

Serverless agent hosting is where Loophole's work lands most directly. Serverless only works economically if cold starts are cheap, and it only works for voice if they are fast. The release also says LiveKit wants to expand beyond voice-only workloads, and AGX-style sandboxes that fork sub-agents in milliseconds point at a broader set of agent jobs, such as tool use and code execution, running on the same substrate.

Flux's read: this is vertical integration by a company that has decided its moat is runtime, not framework. Open-source agent frameworks are easy to copy. A global media network plus a custom hypervisor tuned for bursty agent traffic is not.

What is still unanswered

The announcement is silent on the fate of Loophole's existing products. It does not say whether Architect, a paid product for Kubernetes users, will keep being sold and supported, or whether AGX will still ship to outside customers as its site promised for summer 2026. At least one industry newsletter flagged the lack of any transition notice for current Architect buyers. Acqui-hires of eight-person teams rarely keep standalone products alive for long, and anyone running production workloads on Architect should be asking LiveKit for a written answer.

The other open question is competitive. Sandboxing and fast-boot microVMs have become a crowded category as every agent platform discovers it needs somewhere safe and quick to run code. LiveKit's bet is that doing it in-house, for its own traffic, beats buying it from a sandbox vendor. The under-three-second target is a public number, which makes it a public test. If it lands in the next few weeks as promised, it resets the baseline that other voice-agent platforms will be measured against.