SWE AF

SWE AF is an AI coding agent system built on AgentField for developers delegating changes across a repository or several related codebases. It divides work among agents and adjusts the remaining plan when tasks run into problems.

• Coordinated repository changes: Agents schedule tasks around dependencies and use separate Git worktrees, with support for work spanning an application, shared libraries and dependent services.
• Review and recovery: The classic coding loop uses reviewer and QA feedback to retry work; persistent failures can lead to a different approach, smaller tasks or a revised plan, with explicit technical debt and severity records when scope is reduced.
• Model choice by role: Coding, QA and architecture roles can use different models, with provider support for Anthropic, OpenRouter, OpenAI and Google.
• Resumable builds: Checkpoints let interrupted builds resume, and optional learning passes conventions and failure patterns found during a build to later tasks.

The system runs locally or in Docker and is open source under Apache 2.0. Its model backends require external AI provider credentials and internet access, so self-hosting the agents doesn't make model requests offline.

For work across repositories, failures in the primary application block progress. Failures in supporting dependencies are recorded but don't block the build.