Metatron Research
Engineering knowledge should outlive a coding session.
Metatron Research builds Metatron, an open-source tool that helps software teams give AI coding agents their engineering decisions as versioned repository files. We study how AI coding agents use repository context: who should write it, how agents should find it, and when it helps them do better work.
A repository holds the code a team shipped. It rarely holds all the reasons behind it: why an approach was rejected, which constraint matters, or what a previous failure taught. We build tools and run experiments to make that knowledge available to the next person or agent working in the repository.
Context Inheritance
A Git-Native Architecture and Pre-registered Study of Repository Context for AI Coding Agents
Our paper has been accepted at the 1st International Workshop on Agentic AI for Next-Generation Software Development (AgenticDev 2026), co-located with ASE 2026. We’ll present the work in Munich on October 12, 2026.
In the program: 10:50–11:10 CEST · Session 1: Code Analysis · Forum 8. See the official workshop program for the current schedule.
What we study
The paper introduces the Repository Context Layer: context that is versioned with the code, discoverable by agents, and updated under human review. A pre-registered study on SWE-bench Verified examines how the author of that context, the agent reading it, and its delivery affect the outcome.
Our focus is practical: preserve useful engineering knowledge, give agents a clear way to consult it, and measure whether it changes their behavior. The paper reports the findings and their limits; the public artifact contains the protocol, experiment code, and results.
From research to a working tool
Metatron puts repository decisions into readable Markdown files, with a rule and its rationale. Agents consult those files before changing code and propose updates for review. Teams can use the files directly through Git or add an optional MCP serving layer.
The research studies the underlying context pattern. Metatron is our open-source implementation of that approach. You can inspect both, try the workflow, and see which decisions belong in your own repository.
Watch the checkout demo to see a stored retry rule reach an agent’s implementation and regression test, or set up your first decision.
Who Metatron is for
Metatron is for software teams using coding agents in existing repositories: maintainers who keep explaining the same constraints in review, developers onboarding to unfamiliar code, and teams that want decisions to survive across agent sessions. Researchers can also use the public experiment artifact to inspect and reproduce the study.
The team behind Metatron Research
- Pavel Kerbelpavelkerbel.com
- Milana Kerbel
- Vitali Abramov
- Liat Abramov
Get in touch
Interested in repository context, a replication, or collaborating on the next experiment? Write to hello@getmetatron.com.