Peter Albert is the co-founder of Zeta Labs, the company behind Viktor, an AI coworker for Slack and Microsoft Teams. A former Meta AI engineer and co-author of Llama 2, he builds software that turns language-model capabilities into reliable workplace action.
From foundation models to workplace agents
At Meta, Albert worked on language-model pre-training and post-training and contributed to the Llama 2 research paper. In 2023, he founded Zeta Labs with fellow former Meta engineer Fryderyk Wiatrowski. Their first product, Jace, began as an autonomous browser agent powered by the AWA-1 model before evolving into an AI executive assistant for email, meeting coordination, and related administrative work.
Albert subsequently focused on product and engineering for Viktor, which extends that approach into shared workplace environments. The system integrates with business applications, prepares reports, works with code, and handles operational tasks while requiring approval for sensitive or irreversible actions. In 2026, the company raised a $75 million Series A led by Accel.
How Albert builds useful agents
Keep agents within their training distribution. Albert favors familiar formats such as JSON, XML, and Markdown, concise relevant context, and interfaces that resemble established programming patterns. He breaks difficult decisions into manageable steps instead of making a model resolve multiple unfamiliar abstractions simultaneously.
Add architectural complexity deliberately. His progression starts with prompting, then introduces explicit state tracking, planning, verification, scratchpads, or parallel execution only when the task demands them. Additional components reduce cognitive load but create new failure modes; fine-tuning and reinforcement learning also slow iteration and complicate portability.
Evaluate against realistic work. Albert recommends establishing small, repeatable evaluation sets early and generating synthetic training data from simulated user interactions when fine-tuning becomes worthwhile. Execution feedback and language-model judges can filter unsuccessful examples, while distillation can reduce latency when its performance tradeoffs are acceptable.
Albert’s practical agent-engineering framework links foundation-model behavior to concrete product decisions: use familiar interfaces, measure outcomes, limit unnecessary complexity, and preserve human control over consequential actions.
