Katharina Dzialets

At JetBrains, we used preconfigured environments to delegate feature-sized work on the IntelliJ monorepo. What began as a setup for a few AI champions was soon adopted by entire teams, exposing how difficult it was to share and scale.

Each task started from a hand-written prompt that rarely carried all the necessary context. A generated change could pass every formal check and still be conceptually different from what the team needed, while increasingly detailed prompts only made the agent’s behavior harder to predict. In practice, teams got more value from handing off smaller, repetitive steps. Instead of making each agent run larger and more autonomous, we needed to make the work smaller, more predictable, and reusable.

That led us to build a platform where teams compose larger workflows from shared automations with predefined context, permissions, tools, and network access. We reused the preconfigured monorepo environments and secure proxy layer from the earlier setup. Today, more than 100 teams use the platform, with 935 automations across the SDLC.

In this talk, I’ll explain why delegating entire features did not work well for teams, how smaller shared automations made delegation more predictable, and what 935 automations reveal about the work JetBrains teams choose to delegate to coding agents.