Hailong Zhang is a software engineer and AI product builder behind BabelGPT, Babel Agent, and coding-agent development at Gru.ai. His work focuses on giving autonomous systems specific engineering responsibilities, repository-level context, and human-reviewed ways to contribute production code.
By 2023, Zhang was chief executive of Babel, developing software that translated natural-language requirements into working applications. He argued that probabilistic language models still need deterministic software: models interpret ambiguous intentions, while conventional programs provide reliable computation, structured data, and actionable tools. BabelGPT put that division into practice through controlled, multistep application-generation workflows, combining unstated user context with outputs other software could process.
In 2024, Babel Agent extended those experiments into task decomposition, coding, debugging, documentation research, and testing, escalating to people when requirements or execution became uncertain. Zhang subsequently distinguished product-market fit from product-model fit: successful coding products need both genuine customer demand and models capable of meeting practical quality and speed requirements.
Building agents that earn their autonomy
- Asynchronous coding agents: Unlike IDE copilots that respond while developers type, Zhang’s agents complete bounded repository tasks independently and return proposed changes for human review. Their intended responsibilities include testing, documentation, bug fixes, and code review.
- Test Gru: This agent inspects pull-request changes, identifies missing or outdated unit tests, writes and executes replacements, and submits its own pull request. In 2025, Zhang reported that people accepted more than half its proposed pull requests and that it generated approximately 80 percent of the tests in its own repository.
- Task-specific evaluation and repository context: Zhang builds evaluation datasets and harnesses around concrete jobs, routes workflow stages to suitable models, and described fine-tuning GPT-4o on human-labeled unit tests. His agents assemble relevant context from code, issues, reviews, commits, pull requests, documentation, and framework conventions.
- Agent OS: Shared runtime, tooling, context, and orchestration infrastructure supports additional engineering agents, including systems for refactoring and end-to-end testing.
At AI Engineer Summit 2025, Zhang outlined a division of labor in which agents handle repetitive implementation while people retain architectural decisions, product direction, difficult problems, and final approval.
