What is it about?

Programmers use many separate tools to plan, code, test, and ship software. Each tool has its own window and rules, so people spend time switching between them and stitching results together. This paper explores a simple alternative: place one assistant inside the code editor that listens to the person, understands the task, and then drives the right tools on their behalf.

Featured Image

Why is it important?

This work proposes a single mediator agent inside the IDE that coordinates built-in tools, AI assistants, and external multi-agent systems through one interface. It is timely because developers are already using many separate AI add-ons, which creates friction, context loss, and governance risks. Our contribution is unique in three ways. First, we adapt well known automation levels to software engineering tasks to set clear expectations for responsibility transfer from humans to agents. Second, we define a practical work cycle for the agent that ties user interface, observation, thought, and action to the IDE’s real capabilities, including an explicit action space that spans editor features, build and test, and delegation to external agents. Third, we treat prompts as first-class project artifacts and outline how an IDE can manage them across the software lifecycle. Together, these choices make evaluation possible on real projects and open a path to measurable gains in productivity and reliability while keeping human oversight in place.

Perspectives

I wrote this because I keep seeing developers juggle three or four AI panels while also driving tests, commits, and reviews. The orchestration tax is real. A mediator pattern felt like the simplest way to give people one place to talk, plan, and act without learning every tool’s quirks. Working with an industrial partner sharpened my focus on evaluation inside the IDE rather than in toy sandboxes. My hope is that this paper helps shift the conversation from individual AI features to an interaction model that actually reduces cognitive load. If it nudges teams to treat prompts and agent decisions as shared, reviewable artifacts, that will already be a win.

Ziyou Li
Technische Universiteit Delft

Read the Original

This page is a summary of: Enhancing Human-IDE Interaction in the SDLC using LLM-based Mediator Agents, June 2025, ACM (Association for Computing Machinery),
DOI: 10.1145/3696630.3728721.
You can read the full text:

Read

Contributors

The following have contributed to this page