Mahmoud Mabrouk

Mahmoud Mabrouk is the co-founder and chief executive of Agenta, an open-source platform for building, evaluating, and improving AI applications and agents. His central concern is making AI systems reflect the judgment of the domain experts who understand whether they actually work.

Mabrouk began in academic machine learning, working on computational biology and protein-structure prediction, and earned a doctorate in computer science at Technische Universität Berlin. He subsequently joined a startup working with foundation models, where conventional machine-learning operations failed to accommodate changing prompts, expert evaluation, and production feedback. As he recalled about Agenta’s early years, he pursued an integrated, open-source platform despite investor skepticism about both its breadth and commercial model.

That platform grew to combine prompt management, evaluation, and open-source observability. In July 2026, Agenta 2.0 expanded into a collaborative agent workspace incorporating persistent context, feedback, permissions, human approval, and portable components such as tools, skills, and MCP servers.

• Human-calibrated AI evaluation. Mabrouk maintains that automated judges accelerate development only when their decisions match informed human judgment. In his AI Engineer workshop, he separated an airline-support agent’s failures into policy compliance, response style, information delivery, and tool execution, favoring specialized binary judges with explanatory rationales over generic quality scores.

• GEPA-based reflective prompt optimization. His public workshop materials explore improving evaluator prompts through mutation, merging, and Pareto-frontier selection. He emphasizes that results depend on annotation quality, seed prompts, reflection instructions, model choice, and inference costs; collectively covering every training example does not guarantee a single dependable evaluator.

• Production-data improvement flywheels. Mabrouk connects production traces and expert feedback to error analysis, targeted test cases, revised prompts, and repeated evaluation. He also advocates evaluation-gated prompt deployment, treating prompts as versioned production artifacts subject to testing, approvals, and monitoring.