Vlad Luzin

Vlad Luzin co-founded BAND with Arick Goomanovsky and has worked as its CTO on infrastructure for AI agents to discover one another, exchange context, and collaborate across frameworks and organizations. His work addresses the coordination problems that arise when individually useful agents share a task: preserving history, controlling participation, and keeping people able to intervene.

From multi-agent research to production AI

Luzin’s work on collaborating agents predates the current generation of language models. He led a multi-agent AI incubation team at Samsung Telecom Research, subsequently held product and engineering leadership roles at CME Group, and became vice president of platform services and AI at Verint. At Verint, he led global product, architecture, and engineering for cloud platform services. He holds a bachelor’s degree in mathematics and computer science from Bar-Ilan University and a business master’s degree from the Hebrew University of Jerusalem. Career background.

In November 2022, Luzin joined Lynceus as vice president of product and engineering. Lynceus applied machine learning to semiconductor manufacturing, using fab data to predict problems and help improve production. His responsibilities connected data science, product management, and engineering: developing models, deploying the platform to semiconductor customers, and maintaining those models in production. The role extended beyond making predictions to putting them to use in manufacturing operations.

Luzin also explored software autonomy through Open Elixir Self-Healing Intelligence, a project listed on his professional profile. The FluffyTrain prototype paired two GPT-4-powered applications: one generated and checked Elixir code; the other caught exceptions, supplied the relevant source and error details to the first, and applied the resulting correction through hot code reloading. This made a repair loop concrete: observe a failure, propose a correction, validate it, and update the running code. The project also acknowledged that language models could fail on complex requests.

BAND extends the question of software autonomy to collaboration between agents. Its architecture separates a collaboration mesh, where agents discover peers and delegate work, from a control plane that governs their interactions. Luzin and co-founder Arick Goomanovsky, BAND’s CEO, have developed infrastructure intended to make those interactions possible while controlling who can participate and under what conditions.

Making collaboration reliable

Luzin treats agent coordination as a distributed systems problem. In his argument about agent-to-agent collaboration, a developer running Claude and Codex side by side becomes the router, manually passing messages between stateful participants. A script can automate that exchange, but developers still need to handle discovery, delivery, persistence, and coordination across runtimes. Larger context windows do not resolve those interaction problems.

His writing develops several complementary requirements:

• Reliable delivery and recovery. Routing a message is only the beginning. A system needs to track delivery, distinguish an agent that is still processing from one that has crashed, and recover unfinished exchanges when a participant returns. His production architecture argument puts routing, recovery, loop controls, discovery, and framework interoperability alongside the capabilities of individual agents.

• Shared conversation identity. Frameworks preserve state in different forms, including checkpoints, conversation records, flow objects, and message histories. A handoff can lose the meaning of where the work stands even when both agents successfully exchange text. Luzin’s proposed shared conversation identity gives participants a common history without requiring every pair of frameworks to maintain its own mapping.

• Ownership and permission. Evaluating an agent’s answer does not establish who authorized it, which team owns it, or whether another organization consented to the interaction. His approach to agent governance therefore includes registration, access, visibility, sharing rules, and consent alongside tracing and evaluation.

Keeping the work visible

In multi-agent coding, Luzin identifies another constraint: the developer’s attention. Several agents can produce plans, updates, and review findings at once, while copying a plan into separate sessions creates versions that drift. His writing about Jam and BAND advocates a shared work record containing ownership, blockers, decisions, and progress, with conversations available to explain the work. BAND carries exchanges between participants; Jam makes the work and pending human decisions visible.

This concern connects his experiments in self-repair with his work on collaboration infrastructure. Autonomous software needs more than the ability to produce a response: a repair must be checked, collaborating agents must retain a common account of the task, and participation must be authorized. Luzin’s work makes those requirements explicit and seeks to provide reusable infrastructure for meeting them.