AI agent ShakeFix fixes a pharmacy app bug in 18 minutes at GitLab hackathon
GitLab announced the results of its Transcend hackathon, where participants trained AI agents to take code all the way to production. The winner, ShakeFix, works like this: a pharmacy app user shakes their phone and describes the failure, and four agents write a test, fix the bug and ship the fix in 18 minutes 34 seconds with no humans involved. The takeaway for teams: give agents the reading and editing of code, and leave admission to production to tests and CI.
AI-processed from GitLab Blog; edited by Hamidun News
Writing code with AI is now easy, but getting it from "works on my machine" to production is still painful: tests, security checks, deployment, rollback, incident analysis. GitLab handed exactly this part to the community at the Transcend hackathon, themed "Life after code." In October, participants built agents and flows (chains of steps executed by AI agents) on the GitLab Duo Agent Platform and decided how much freedom to give them, from human approval of every step to full autonomy.
The most vivid example is ShakeFix. A user of a pharmacy app shakes their phone and writes what broke, and four agents write a failing test, fix the bug and ship the fix on their own. In a live run, this took 18 minutes 34 seconds with no humans involved.
In total, more than 3,300 developers registered and 783 projects were submitted, with a $45,000 prize pool across nine categories.
The reason to read about this is the design pattern, not the prizes: the winners drew a clear line between what the agent does, what automation verifies and where a human signature is needed.
Path A: projects from scratch
Aftermath (Ahadi Cyizere, best assisted agent, $4000) answers a question that scanners do not cover. When a compromised npm package is disclosed, you need to know whether it managed to run in CI last week. A flow of three subagents reconstructs past CI jobs and the list of secrets each one could reach. A human approves the scope of the investigation, then the containment plan, and closes the incident after CI has replaced only the compromised keys and deleted the affected images.
Lookback (Ujwal Suresh Vanjare and Arpita Madhukar Kalburgi, best supervised agent, $4000) starts from a successful appeal. If benefits-payment software applied the rules incorrectly to one person, the flow analyzes the decision, reproduces the error with a failing test and opens a merge request with a fix. After the merge, ordinary code re-runs past decisions and finds everyone else affected. In the demo, it found 413 underpaid families, and the amounts owed were published as a package for the agency.
ShakeFix (Venkatesh Devale, best autonomous agent, $5000) compares the fix against production behavior and only then merges and ships the code. That is what turned a phone shake into a working fix in 18 minutes 34 seconds.
Path B: existing open-source projects
Unstuck (Aryan Shah and Riken Shah, best assisted agent, $4000) finishes the dependency updates that bots start but do not complete. A bot bumps the version number but cannot read release notes or fix what breaks. Two flows plan an update campaign and carry it through a single merge request, citing the release note under each change and asking a human when behavior changes. On the doccano annotation tool, which was stuck on an unsupported Django with 80 known vulnerabilities, this resulted in zero vulnerabilities and Django 5.2 across six verified merge requests.
CodeTinder (Jason Yu, best supervised agent, $4000) starts from the idea of "review less code, teach agents more." The agent picks a function in a merge request and writes several variants that pass the tests. The developer swipes through them on a phone, Tinder-style. The swipes update a versioned principles document in the repository. All agents read it, then rewrite the rest of the merge request to match the new preferences. The project also won the "Most Creative" prize ($4000): it was the highest-rated idea across both paths.
Locus (Ahammad Shawki, Nourin Richi, Mohammad Ashfaqul Hasan, best autonomous agent, $5000), built on the JSCAD library, turns a math teacher's email describing a function into a tactile graph for a blind student. The agents draw a graph with raised lines and Braille labels, check it against research-based tactile reading rules, fix their own mistakes, publish the result to Google Cloud, send the print files to the teacher and order the print. The agents cannot mark the work as finished themselves: the graph is released only if it passes the rules.
Special prizes and takeaways
The "Most Stages" prize goes to projects that touch the most of the nine stages of the DevSecOps lifecycle. One project won on each path, $5000 each.
Stale (Diptarka Nandi, Harshit Anand, Abhinav Kumar, Ashutosh Pandey) re-checks the "safe to ship" verdict when the world changes, not just the code. For example, when a root certificate is no longer considered trusted. A scheduled pipeline watches external sources and creates a task for a flow that writes the fix. CI confirms it, and a canary release verifies it before traffic is switched.
Earned Autonomy (Shivansh Kalra) assigns each merge request in a FastAPI template its own autonomy level, from "no intervention" to "blocked," based on accumulated evidence. Five agents handle planning, triage, infrastructure review, routing and rollback, and they can only raise the risk score. In one run, a vulnerable dependency went from alert to fix in production in 35 minutes with no humans involved; in another, a bad release was rolled back in 3.3 seconds.
Memory Lane (Oliver McBride, "Greenest," $5000) helps nursing home coordinators prepare reminiscence sessions for people with dementia. Three flows take an ordinary task through a security check, CI/CD and Cloud Run into production. The author measured the carbon footprint of each shipped request using the Software Carbon Intensity specification and rebuilt the pipeline to build once and skip redundant runs. The number of CI runs per shipped request dropped by 48%.
The common pattern GitLab notes: agents read, write and route, while tests and CI decide whether a step has passed, and in most projects a human signs off on risky deployments. A caveat: these are selected hackathon demos, and figures like 18 minutes come from the authors' showcase runs, not from long-term operation. But the pattern transfers: define which steps an agent can do on its own, which are verified by automation and where a human signature is needed.
- Nov 18, 2026rust-anything teaches AI agents to rewrite apps in Rust module by module
- Nov 18, 2026Google finds AI agents that discover vulnerabilities and hack websites on their own
- Nov 18, 2026BioMaster Agent on Kimi Picks 100 Genes in 90 Minutes Instead of 4 Hours
- Nov 18, 2026Hugging Face routes LLM requests through OpenRouter, cutting capacity errors by 46%
Key Questions Answered
What are the key details on Path A: projects from scratch?
Aftermath (Ahadi Cyizere, best assisted agent, $4000) answers a question that scanners do not cover. When a compromised npm package is disclosed, you need to know whether it managed to run in CI last week. A flow of three subagents reconstructs past CI jobs and the list of secrets each one could reach. A human approves the scope of the investigation, then the containment plan, and closes the incident after CI has replaced only the compromised keys and deleted the affected images.
What are the key details on Path B: existing open-source projects?
Unstuck (Aryan Shah and Riken Shah, best assisted agent, $4000) finishes the dependency updates that bots start but do not complete. A bot bumps the version number but cannot read release notes or fix what breaks. Two flows plan an update campaign and carry it through a single merge request, citing the release note under each change and asking a human when behavior changes.
What are the key details on Special prizes and takeaways?
The "Most Stages" prize goes to projects that touch the most of the nine stages of the DevSecOps lifecycle. One project won on each path, $5000 each. Stale (Diptarka Nandi, Harshit Anand, Abhinav Kumar, Ashutosh Pandey) re-checks the "safe to ship" verdict when the world changes, not just the code. For example, when a root certificate is no longer considered trusted. A scheduled pipeline watches external sources and creates a task for a flow that writes the fix.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.