Descriptions:
Robert Brennan, co-founder and CEO of OpenHands (the MIT-licensed coding agent originally launched as OpenDevin), presents a detailed orchestration architecture for using parallel AI agents to automate large-scale software refactors—the kind of tech debt cleanup and code modernization work that is too large for any single agent to handle in one pass.
The central technique involves constructing a dependency graph of the codebase and partitioning it into PR-sized batches using directory structure as a semantic grouping heuristic. Batches with no upstream dependencies can be dispatched in parallel immediately. Each batch is annotated with complexity metrics that signal to human reviewers how much scrutiny a given set of files warrants before an agent’s changes are accepted. A verifier component—either programmatic, running unit tests or linters, or LLM-based, evaluating code against a defined set of smell rules—validates each batch and updates its status, enabling the orchestration system to track progress across the full refactor.
Brennan situates the work in a broader argument: even without further model improvements, the organizational and workflow barriers to AI adoption in software engineering are still falling, meaning orchestration patterns like these will become increasingly mainstream. The talk draws on OpenHands’ real operational experience and includes a live visualization of the dependency graph tooling. Engineers managing large legacy codebases, planning modernization projects, or exploring how to structure multi-agent coding pipelines will find the batching and verification approach directly applicable.
📺 Source: AI Engineer · Published January 08, 2026
🏷️ Format: Hands On Build
