As we hand more autonomy to LLM agents, the hard engineering problem stops being "can the agent do the task" and becomes "how do we keep a human meaningfully in control without making them the bottleneck?" This talk shares a concrete pattern I built and run: a harness of narrow, single-purpose agent "lenses" that detect and propose but are structurally prevented from acting.

Each lens has exactly one job and runs on a schedule — watching a set of repositories, scouting ideas, surfacing opportunities — and emits structured output (cards on a project board) rather than taking action. A planning lens breaks those into small milestones, but only proposes; it never builds. The flow is gated by labels: items needing a human decision route to me, and only a human-applied "ready-to-build" label can trigger the one lens allowed to write code. The human checkpoints are structural, not a matter of discipline: an agent literally cannot promote its own work past a gate.

The pattern is demonstrated on an open-source agent runtime, but it's framework-agnostic — anyone building toward agent autonomy can apply "narrow agents + structural human gates + propose, don't act."
Sobre Nadia Ujovich: Soy ingeniera en sistemas de información recibida en 2017. Tengo más de 12 años de experiencia trabajando con tecnologías backend como Java y Python, pero de a poco empecé a transicionar al desarrollo web. Además, soy profesora de Informática Aplicada en una Licenciatura orientada a Deportes. Aficionada al Data Science y a la Inteligencia Artificial. Show less