Anita Kirkovska is Head of Growth at Vellum, where she helps shape how developers build reliable AI agents and how personal assistants learn their users’ preferences without exceeding their authority. Her work spans agent evaluation, developer tooling, technical education, and assistants with persistent memory.
Originally from Macedonia and based in New York, Kirkovska earned a master’s degree in data science at the University of Central Florida. Before joining Vellum, she ran an AI agency building and testing products with large language models. Her announcement of her move to Vellum highlighted problems she had encountered firsthand: prompt comparison, version control, performance monitoring, and output evaluation. As Vellum’s founding growth lead, she has published extensive technical guidance on prompting, retrieval, model selection, production operations, and agent architecture.
Test-driven development for AI agents. Kirkovska argues that reliability comes from evaluating entire workflows, including intermediate decisions, tool selection, and failure modes. Her AI Engineer Summit talk outlines a cycle of experimentation, domain-expert evaluation, deployment, tracing, regression checks, model fallbacks, and production feedback. Teams must explicitly balance quality, latency, cost, and privacy.
Agent autonomy as a spectrum. She distinguishes ordinary model calls from systems that select tools, plan multistep tasks, maintain context, and eventually act independently. Greater autonomy demands closer scrutiny of memory, intermediate decisions, and stopping conditions.
An evaluable SEO agent. Her open-source SEO agent researches competing articles, identifies coverage gaps, drafts content, and routes it through an editor-like evaluator. Conversational memory carries feedback into bounded revision loops, keeping research, generation, evaluation, and termination inspectable.
Personal intelligence with durable memory. Kirkovska’s personal-assistant product writing describes working-memory files, selectively retrieved knowledge graphs, procedural learning, and prioritized user corrections. Her agent-plugins announcement extends assistants through packaged skills, tools, hooks, and rules, with approval required for sensitive actions.
Kirkovska has also introduced the Vellum Workflows SDK, which synchronizes code-defined agent workflows with a visual interface, and maintains a model-routing project for directing tasks across configurable model providers.
Reliable agents need more than capable models: representative tests, observable decisions, and bounded feedback loops turn promising examples into workflows that can improve in production.
