Context Inheritance, Published in ACM
A coding agent should be able to learn from the knowledge already earned in a repository. Our published paper studies when that works, who should author the context, and what the evidence actually supports.
Our paper, Context Inheritance: A Git-Native Architecture and Pre-registered Study of Repository Context for AI Coding Agents, is now published in the ACM proceedings of AgenticDev ’26. I co-authored it with Milana Kerbel, Vitali Abramov, and Liat Abramov, and presented it in Munich on October 12 at the workshop co-located with ASE.
From a context file to a research question
The starting idea was the Repository Context Layer: keep the reasons behind engineering decisions beside the code, let agents consult them before working, and review proposed lessons through Git. Context.md describes the contract; Metatron implements it.
But storing context is only the beginning. We wanted to know who should write it, which agents benefit from it, and whether the way it reaches the agent matters. The paper pairs the architecture with a pre-registered study on SWE-bench Verified.
The finding I want readers to take away
On tasks sharing relevant project constraints, an 8B local model’s code-localization accuracy rose from 21.1% to 47.2% when it read context written by a frontier model. Its own lessons did not improve that measure. Useful knowledge can come from a more capable author and be reused by a smaller agent: context inheritance.
That result concerns finding the right code. It does not establish better bug-fixing success for the smaller model. Context also failed to improve the frontier agent’s resolve rate on a broad random sample. The paper reports the other results and their statistical limits; the practical lesson is selective, reviewed context, with evidence for where it helps.
Read, inspect, or watch
- Publication page — findings, citation, and links to the ACM version.
- The research article on Metatron — a longer explanation of context inheritance.
- Replication materials — protocol, experiment code, and results.
- Munich presentation — slide viewer and downloadable PDF.
The repository should be able to carry what a team has learned, as well as what it has built. This paper is a first measured step toward that workflow.