OpenHands Reveals the Architecture for Massive Automated Refactoring

"If our jobs are going to change, we want that change to be driven by the software development community." This sentiment, voiced by Robert Brennan, CEO and Co-founder of OpenHands, underscores the philosophy behind automating the most tedious and necessary tasks in software engineering. Brennan’s presentation focused squarely on the limitations of current large language model (LLM) agents and introduced the concept of agent orchestration as the vital next step in achieving true, large-scale automation. While atomic coding tasks are now readily handled by single agents, monumental efforts like migrating entire codebases or resolving company-wide technical debt require a coordinated swarm.

Robert Brennan spoke at the AI Engineer Code Summit about leveraging parallel agents to automate massive refactors, dependency updates, and code modernization tasks—the "toil" that consumes significant developer hours across the industry. OpenHands, an MIT-licensed coding agent project, is designed specifically to tackle these non-atomic challenges, moving beyond simple code generation toward autonomous execution of complex, multi-step engineering projects. The evolution of LLM coding capabilities demonstrates a clear trajectory from simple, context-unaware code snippets in 2022 to the current frontier of agentic orchestration expected to dominate 2025 and beyond.

The initial wave of AI coding assistants, such as GitHub Copilot, provided massive productivity boosts by offering context-aware code generation within the IDE. This shifted the developer’s role from typing lines of code to guiding the agent through short, atomic tasks, like writing a suite of unit tests for a single function. Brennan noted that for himself and his team, this shift has been profound: "The mentality of the job hasn't changed, but what the actual work looks like has changed quite a bit." The mental load of critical thinking remains, but the mechanical execution of coding has been largely outsourced.

However, even the most capable single agents stumble when faced with large, interconnected changes. This failure stems from inherent technological and human limitations. Agents suffer from limited context windows, "laziness" (refusing to tackle large tasks), lack of deep domain knowledge, and compounding errors where a small initial mistake cascades into an unusable output. Humans, conversely, struggle to define the "done" state unambiguously and find it difficult to decompose massive tasks accurately.

The core innovation required is orchestration: coordinating multiple agents, often in parallel, with a continuous human feedback loop. OpenHands aims for a 90% automation, 10% human effort split, where human oversight is focused on high-level strategy and intermediate verification rather than debugging individual lines of code. For instance, one OpenHands client saw a "30x improvement on time to resolution for these CVEs by doing orchestration at scale."

This paradigm shift creates the role of the "Agent Orchestrator." Every developer will utilize agents for day-to-day tasks, but Brennan estimates that only about 5% of developers will specialize in this new discipline. These orchestrators, capable of decomposing large tasks and managing fleets of agents, unlock massive productivity lifts (100% to 3000%) across high-leverage, repeatable tasks.

The key to successful orchestration is task decomposition, which requires breaking down an end-goal into tasks that are one-shot shottable, fit within a single commit, can be executed in parallel, and are verifiable by a human as correct or incorrect. This requires analyzing the codebase's inherent dependency structure. Rather than attempting to migrate an entire application from Redux to Zustand at once, the orchestrator identifies the leaf nodes of the dependency tree—those components that rely on the fewest internal dependencies—and tasks parallel agents to migrate those first.

This dependency-tree approach ensures that changes are introduced sequentially, preventing cascading failures and allowing the human orchestrator to review intermediate, manageable pull requests. For large-scale maintenance, such as resolving CVEs across thousands of repositories, the agent architecture can scan the entire codebase, spin up a separate agent for each vulnerability, and update the dependency, fix breaking API changes, and open a pull request automatically. The resulting pull requests are small, tightly focused, and easily merged, drastically reducing the latency between vulnerability announcement and resolution. This architecture allows enterprises to systematically address modernization efforts, technical debt, and large-scale migrations that were previously considered prohibitive due to their sheer volume and complexity.