Developers become Orchestrators: From Human-in-the-Loop to Spec-in-the-Loop

Development teams still struggle with a fundamental question: what can we safely delegate to AI, and what must remain human-owned? Despite rapid progress in AI tooling, most teams are stuck in “copilot mode” — gaining speed, but also increasing cognitive load, risk, and inconsistency.

In this talk, we show why and how the role of the developer must evolve: from writing code and managing endless shift-left concerns, to designing intent, constraints, and guarantees — and orchestrating AI-driven execution.

We introduce spec-driven development as a governance and execution model for AI-assisted software delivery. Instead of focusing on prompts or tools, we frame AI delegation as a risk- and criticality-aware decision: humans define what must be true (functional specs, architecture decisions, security, SRE, and compliance constraints), and AI is delegated everything else.

To make this concrete, we present a realistic end-to-end experiment: a non-critical internal service built with maximum AI delegation. Humans provide: Functional specifications and ADRs, Non-functional requirements (security, reliability, coding standards) and Release and compliance constraints

From there, AI generates implementation, tests, CI/CD pipelines, packaging, and deployment artifacts — fully automated.

On stage, we demonstrate how changing a specification (for example an SLO or security requirement) triggers repeatable, auditable regeneration of code and infrastructure — without manual re-implementation. We deliberately push this approach to its limits to show where it works, where it breaks, and why it must never be applied blindly to critical systems.

Attendees will leave with:

A clear understanding of the guardrails required for safe AI delegation. A practical delegation matrix for AI-assisted development. A mental model for specifications as an execution boundary, not documentation. A realistic, experience-based view of fully automated delivery for low- and medium-critical systems.