Nikolay Petrov

Nikolay Petrov contributes to Dokimi, Intel’s open-source toolkit for testing how written instructions affect coding agents. He studies two questions that matter when editing a skill file. What can the model already explain without the instructions? Which instructions does the agent still need to carry out the task?

In 2023, as an AI software engineering manager at Intel, Nikolay worked with Dmitry Razdoburdin on a different performance question. They showed how developers could keep their training code and get faster CPU predictions from a trained XGBoost, LightGBM, or CatBoost model. Using daal4py, the Python interface to Intel’s oneDAL library, developers could convert a model from any of the three frameworks and call predict(). In their benchmarks, the converted models produced predictions faster without losing accuracy.

Nikolay’s AgentReady scanner examines what a coding agent can discover when it enters a repository. For example, a project may pass its continuous integration checks yet give an agent no clear command for running the tests locally. AgentReady can flag the missing command and compare two revisions so a maintainer can catch the omission during review. It reads files on disk without executing the project’s scripts. Even when scanning many repositories, it works from local copies, so teams can use it without giving it credentials for their GitHub organization.

Which skill instructions earn their place?

In November 2026, Nikolay used Dokimi to test 478 facts from 15 Superpowers skills. He compared answers to the same question with and without the relevant skill. On test-driven development, Haiku 5.5 and Sonnet 5.5 each answered 24 of 25 questions correctly without the file when the grader accepted the general idea. Elsewhere in the collection, exact filenames and required formats were often what the skill added. Nikolay recommended preserving those details and trying cuts to familiar explanations. A reminder can still steer the order in which an agent works, so his test for keeping a cut is to rerun the coding tasks and check whether the agent still follows the intended procedure.

Nikolay also proposed a placebo test for skill authors. In experiments on a collection of 33 Intel skills, an unrelated passage the same length as a skill reduced task scores for some models. Including that comparison helps distinguish the effect of the instructions from the effect of adding text at all. He also called for testing each supported model separately. One skill improved performance for several models, yet led another to announce what it planned to look up and stop before answering. A pooled score could hide the failure that users of that model would encounter. Nikolay’s proposed changes to the Agent Skills evaluation guide