Agent-Field/SWE-AF

What it solves

SWE-AF addresses the limitations of single-agent coding loops by providing a coordinated, factory-style approach to software engineering. It automates complex, multi-issue programs that require planning, execution, and governance, moving beyond simple code generation to full end-to-end software development.

How it works

The system uses a coordinated control stack of specialized agents (Product Managers, Architects, Coders, Reviewers, and Testers) organized into three nested control loops:

• Inner loop: Handles single-issue retries when QA or reviews fail.
• Middle loop: Uses an advisor to split work or re-approach issues when the inner loop is exhausted.
• Outer loop: Triggers a replanner to restructure the remaining task graph (DAG) when escalations occur.

It supports single-repository and multi-repository modes, allowing it to manage changes across primary applications and their dependencies simultaneously.

Who it’s for

• Software Engineering Teams: To automate repetitive refactoring, feature implementation, or bug fixing.
• DevOps/Platform Engineers: To scale autonomous software engineering through fleet-scale orchestration.
• Developers seeking high-autonomy tools: Who need a system capable of managing hundreds to thousands of agent invocations for complex builds.

Highlights

• Multi-model orchestration: Assign different LLMs to different roles (e.g., a high-reasoning model for architecture and a faster model for coding).
• Hardness-aware execution: Automatically adapts task depth and replanning based on the difficulty of the issue.
• Multi-repo support: Orchestrates coordinated changes across multiple codebases and microservices.
• Continual learning: Injects patterns and conventions discovered during a build into downstream tasks.
• Checkpointing: Supports resuming builds after crashes or interruptions.