Automating Large Scale Refactors with Parallel Agents - Robert Brennan, AllHands

Four distinct phases of AI coding

The field progressed from context-unaware snippets to IDE-integrated autocomplete (GitHub Copilot), then to autonomous agents capable of running code and debugging (OpenHands/Devin), and now to parallel orchestration where multiple agents coordinate on complex tasks.

Autonomous agents mark a step-change

Early 2024 introduced agents that could execute code, search error messages, and iterate independently, automating the entire inner loop of development rather than just generating text.

Orchestration is the bleeding edge

Current top-tier adopters are running cloud-based agents in parallel with sandboxed environments, allowing agents to spawn sub-agents and tackle enterprise-scale refactoring safely.

CVE remediation at massive scale

One client with tens of thousands of developers across thousands of repositories achieved 30x faster vulnerability resolution by orchestrating agents to scan repos, update dependencies, and open PRs automatically.

Systematic code modernization

Effective for tedious but complex migrations like Spark 2 to Spark 3, adding Python type annotations, or migrating React Redux to Zustand—tasks too large for single-agent completion.

Proactive maintenance from logs

Organizations are using agents to monitor error logs, detect new patterns, and automatically submit patches to add error handling before incidents escalate.

Context window limitations

Large-scale refactors spanning extensive codebases exceed LLM context limits, forcing context compression that causes agents to lose track of architectural patterns.

Error compounding over long trajectories

Small initial mistakes multiply when agents execute hundreds of steps, and phenomena like 'laziness' cause agents to quit after partial completion, claiming human teams are needed.

Missing domain intuition

Agents lack the implicit mental models engineers hold about their specific architectures, making naive attempts at tasks like monolith-to-microservices decomposition without human guidance ineffective.

Decomposition is critical

Engineers must break massive projects into discrete, verifiable sub-tasks that individual agents can execute independently before aggregating results into cohesive changes.

Intermediate review checkpoints

Rather than aiming for 100% automation, successful workflows target 90% automation with mandatory human review between steps to catch errors before they compound.

Asymmetric productivity gains

While most developers using single agents see ~20% productivity lifts, the top 1% using orchestration for specific tech debt categories achieve 30x (3000%) improvements, completing years of backlog in weeks.