If you run Loophole Labs' Architect on your Kubernetes clusters, the company that sells it to you now works for a voice-AI platform, and nobody has told you what happens to the product. LiveKit announced on September 24, 2026 that it has acquired Loophole Labs, that all eight of its engineers are joining, and that it wants Loophole's technology to get agent startup times under three seconds. The announcement does not mention Architect or AGX, the two products Loophole sells or is preparing to sell. Architect customers need a written continuity and notice commitment before the integration work pulls the team away. LiveKit Cloud buyers should hold off on repricing concurrency until the under-three-second claim is actually running in production.
The deal is a clean fit for LiveKit. For Loophole's own customers, it is an open question.
What Did LiveKit Actually Buy?
LiveKit bought a hypervisor and the team that built it, not a product line. The press release describes Loophole as having "developed its own virtualization technology, including a hypervisor called Substrate, to enable workloads to run on and move between machines with little to no disruption." Terms were not disclosed. Founder and CEO Shivansh Vij keeps leading the team, and the release says LiveKit "plans to bring Loophole Labs' technology into production over the next several weeks."
LiveKit's stated reasons are all about its own operations: start agents faster, respond to demand spikes, "reduce operational costs because agents no longer have to sit idle," and "move closer to offering customers essentially unlimited concurrent agents." CEO Russ d'Sa said the technology "fills an important gap for us as we work to become the default platform for building and running agents, including workloads that go beyond voice." Vij said the team had been "working closely with LiveKit over the last year," so this is a partnership that became an acquisition. The integration was already planned.
LiveKit can afford the move. It raised $100 million at a $1 billion valuation in January 2026, led by Index Ventures with Salesforce Ventures participating, and it names Salesforce, Meta, Microsoft, Spotify and 911 emergency services as customers in the acquisition release.
What Happens to Architect and AGX?
Nobody has said, and that silence is the risk. As of September 24, Loophole's homepage still markets both products and carries no transition notice. Its blog has nothing on LiveKit either.
The two products are at very different stages, and the exposure is different for each.
Architect is a paid product with published list prices. It checkpoints running Kubernetes pods (memory, open sockets, in-flight work) before a spot instance is reclaimed and restores them on a fresh node. The vendor claims "~10s live migration," "0 data loss · 0 dropped connections," sub-50ms wake from hibernation, and savings of "up to 90%" against on-demand. Support is narrow today: Amazon EKS on AL2023, Kubernetes 1.33+, two nodes minimum. Pricing runs from $1,500 a month for 10 nodes (Founder) to $3,500 for 20 (Team), $10,000 for 50 (Fleet), and custom Enterprise with "annual and committed pricing available." The Fleet tier promises a one-hour response on critical issues, 24/7.
A one-hour critical SLA is a staffing promise. Eight engineers are about to spend "the next several weeks" putting Substrate into LiveKit's production stack. If your cluster is running stateful workloads on spot because Architect told you it was safe, the question is who answers the page during that integration sprint, and for how long after it.
AGX is not generally available. Its site describes sandboxes that boot a 2 GiB microVM in 187ms and fork in 7ms, says it is "Coming Summer 2026," and asks visitors to join a waitlist or become a design partner. If you are a design partner, you have a prototype and a roadmap conversation with a team that now has a different employer. Treat that roadmap as void until LiveKit restates it.
Has Loophole Pulled Back From a Product Before?
Yes, and that's the most useful precedent here. Loophole's earlier open-source live-migration project, Drafter, described as "a compute primitive designed for live migration" and built on a custom Firecracker fork under AGPL-3.0, was archived on October 15, 2025 and is now read-only. Substrate, Loophole's in-house hypervisor, is the asset LiveKit just bought, and nothing public says it is open source. The company's public GitHub repositories are networking and tooling libraries, not the hypervisor.
That is a normal choice for a startup. It also tells an Architect customer something: when the core technology moved on, the older artifact was frozen rather than handed to a foundation. Nobody should assume a product outside the acquirer's core business gets a different treatment.
The pattern is familiar from this year. Baseten's purchase of Blaxel came with a "keeps running" pledge and no end date. Klaviyo's Agency deal ended the product within weeks. Harvey's Guardrails AI acqui-hire sunset the Hub. LiveKit has so far made no promise at all, not even a vague one.
Why Does a Three-Second Cold Start Matter to LiveKit Buyers?
It matters because idle capacity is what you pay for today, one way or another. A cold start is the delay while a new agent process spins up and loads its models before it can join a call. LiveKit's docs say the framework runs each agent job in its own process and keeps a pool of prewarmed idle processes to hide that delay: by default math.ceil(cpu_count) in Python, capped at four in Node.js, and zero in development mode.
On LiveKit Cloud, the cost of that pool shows up in plan limits. LiveKit's own blog post on join latency says agents on the Build plan "scale down to zero" once sessions end, "so the next join has to start from cold." Cold start prevention is listed as a feature of the $500-a-month Scale tier, not the $50 Ship tier. Scale caps you at "up to 600" concurrent agent sessions, and Enterprise is custom-priced.
"Essentially unlimited concurrent agents" is aimed straight at that ceiling. If LiveKit can bring an agent up in under three seconds from nothing, it doesn't have to keep a warm pool per customer, and a concurrency cap turns into a policy choice rather than a capacity limit. That would be good for buyers, but it isn't shipping yet. The release says "several weeks," and a three-second start is still three seconds of dead air on a phone call if it lands on the first ring.
The steel-man for moving fast: LiveKit and Loophole have worked together for a year, so Substrate may be closer to production than "several weeks" suggests, and a buyer who signs a large Enterprise concurrency commitment now could lock in a price just before the cost basis drops. That argument works for LiveKit. For you it's a reason to keep the term short, not to sign a long one.
Does Substrate Reach Self-Hosted LiveKit?
The release doesn't say, and for many enterprises that is the bigger question. LiveKit Agents is open source, downloaded "over 1M times a month" per the company, and many contact centers self-host it next to their own telephony. The case for self-hosting is cost: owning more of the voice stack is how teams cut spend. If Substrate stays Cloud-only, the fastest agent startup becomes a reason to move onto LiveKit's meter, and self-hosters keep paying for idle processes on their own nodes.
Neither answer is wrong. You just need to know which one you're planning around before you set next year's infrastructure budget.
Will Your Data Processing Terms Change?
Probably not in a way that triggers anything automatically. For LiveKit Cloud customers, the relevant clause is the sub-processor one. LiveKit's data processing addendum commits to "no less than ten (10) days" written notice before adding or replacing a sub-processor, with an objection right "based on reasonable grounds relating to data protection." Loophole's engineers joining as employees doesn't need a notice. If Substrate runs on infrastructure LiveKit doesn't already use, a new sub-processor could. Subscribe to the list.
For Architect customers, the question runs the other way. Your contract is with Loophole Labs. Whether it now sits inside LiveKit, and on what terms, is exactly what the announcement leaves out.
What Should You Do About It?
The work splits by which side of the deal you're on.
This Week:
- Architect customers: ask for the continuity statement in writing. Email your Loophole contact and ask four things: will Architect be sold and supported after the integration, for how long, will the Fleet and Enterprise SLA response times hold during it, and how many days' notice you get before end-of-sale or end-of-support. "Keeps running" with no date is not an answer. Ask for 12 months' notice.
- Architect customers: list which workloads rely on live migration. Anything stateful that you moved to spot because of Architect falls back to the two-minute interruption notice AWS gives by default if Architect goes away. Know which of those you'd move back to on-demand, and what that costs. Our GPU and spot pricing breakdown has the math.
- AGX design partners: freeze integration work. Don't build further against a pre-GA API whose roadmap owner just changed. Alternatives with public pricing exist, including E2B, Daytona and Modal.
This Month:
- LiveKit Cloud buyers: measure your own cold starts now. Log agent join latency for a week at current settings. You need a baseline to hold LiveKit to the under-three-second claim when it ships.
- LiveKit Cloud buyers: keep any concurrency commitment short. If you're negotiating past the Scale tier's 600-session cap, ask for a 12-month term with a repricing clause tied to the Substrate rollout, not a multi-year fixed commitment.
- Self-hosters: ask LiveKit directly whether Substrate lands in open-source Agents, in a self-hosted enterprise build, or in Cloud only. Put the answer in your 2027 infrastructure plan.
Before Renewal:
- Add a change-of-control notice clause to every agent-infrastructure contract. Blaxel, Loophole and the rest of the agent-runtime layer are being bought one at a time. The clause costs nothing to ask for and is the only thing that turns silence like this into a deadline.
The Bottom Line
This is a good deal for LiveKit and an uncomfortable one for Loophole's customers. LiveKit bought a hypervisor to fix its own biggest cost: idle agents waiting for calls that haven't arrived. Architect's spot-migration business and AGX's sandboxes don't obviously fit a company that describes itself as a platform for voice, video and physical AI agents, and the announcement didn't try to make them fit.
Acquirers usually find out what they bought after the deal closes. As a customer, you should find out before they do.
Ask for the date while they still have time to answer.
Continue Reading
- Baseten Buys Blaxel, Whose 'Keeps Running' Pledge Has No End Date
- Eleven Voice Agents, One Bank Call, No Clean Winner
- Best AI Contact Center Platforms: Containment Isn't Resolution
- Stilla Promised Continuity. Meta Bought It for WhatsApp.
- Run:ai vs Kubernetes vs Slurm: Static Quotas Strand Your GPUs
- Klaviyo Bought Agency. The Product Dies August 31.
- Upwind Buys Aegis, a Nine-Month-Old Startup With No Customers