AI Coding Agents Take Center Stage: JetBrains × Codex Hackathon Winners Revealed

Breaking News — The winners of the first-ever JetBrains × Codex Hackathon have been announced after a weekend of intense development. Over 40 teams competed to build AI-native tools directly inside the IDE, with hyperreasoning taking first prize for its breakthrough approach to coding agents that don't just call an LLM once but search through potential solutions.

Winners at a Glance

First Place: hyperreasoning — Aditya Mangalampalli
Second Place: Scopecreep — Bhavik Sheoran, Kenneth Ross, Roman Javadyan, Joon Im
Third Place: mesh-code — Ayush Ojha, Coco Cao, Kush Ise, AL DRAM

Three additional finalists were also recognized for their innovative projects that integrate AI deeply into the development workflow.

AI Coding Agents Take Center Stage: JetBrains × Codex Hackathon Winners Revealed — Source: blog.jetbrains.com

First Place: hyperreasoning — Solving the 'Thinking in Circles' Problem

Most coding agents call a model once and hope for the best. Hyperreasoning replaces that single shot with a search process: the system drafts several possible approaches, then a learned controller decides which to expand, which to cut, and which to verify against tests.

“LLMs spend a lot of time thinking in circles,” said creator Aditya Mangalampalli. “Our controller feeds compiler errors and failing tests back into the decision loop, so the agent learns on the fly.”

Inside the IDE, a tool window renders the search live. Developers watch which paths the controller explored before settling on one. The project argues that a smaller local model wrapped in this verified search loop can compete with far larger frontier models at lower cost — making reasoning visible and directable inside the editor.

Second Place: Scopecreep — Hardware Bring-Up in One Window

Hardware debugging typically juggles schematic viewers, vendor apps, a terminal, and a spreadsheet. Scopecreep collapses that into a single JetBrains tool window. Hand it a circuit schematic and an agent walks through testing the board — picking signals to measure, capturing readings, and producing a report.

The key design choice: when the agent decides a probe must be placed, the session pauses and shows the engineer exactly where to put it. The engineer places the probe physically and clicks Resume. It’s a smart human-in-the-loop approach for real instruments on a real bench.

Third Place: mesh-code — Agent Memory That Persists Across Machines

Switch laptops mid-task and most coding agents start from scratch. mesh-code gives agents shared memory of an in-progress project: what’s been tried, what’s been decided, what’s still pending. A session beginning on one laptop can continue from another, with whichever agent is available. Codex is one of the agents that can plug into this memory layer.

Background: The Hackathon’s Vision

The JetBrains × Codex Hackathon took place over a single weekend, with roughly 40 submissions from teams around the world. The goal was to build AI features that are native to the IDE, not bolted on top of it. Participants explored what it means to direct an agent, watch how it reasons, manage its attention, and decide when its output is ready to ship.

Organizers emphasized that the IDE is evolving from a place where developers write code into a place where they collaborate with an AI agent. The winning projects exemplify this shift, embedding transparency and control directly into the development environment.

What This Means for Developers

The hackathon results signal a clear direction: AI coding agents are moving beyond simple autocomplete. Hyperreasoning shows that search-based reasoning can achieve frontier-level performance with smaller models. Scopecreep demonstrates that autonomous agents can still respect human expertise where physical safety matters. Mesh-code hints at a future where agent state is portable and persistent across workspaces.

For developers, this means the IDE is becoming a command center for AI-powered workflows. The winners show that the best agents don't just generate code — they explain their reasoning, accept direction, and integrate with real-world hardware and team collaboration. As these tools mature, they promise to drastically cut debugging time and make AI assistance more reliable and transparent.

This is a developing story. Stay tuned for more details on the remaining finalists and future hackathon plans.