Agent-Field/SWE-AF

What it solves

SWE-AF is an autonomous software engineering factory that allows users to ship complex software changes end-to-end via a single API call. It moves beyond simple single-agent coder loops by coordinating a full virtual engineering team—including product managers, architects, coders, reviewers, and testers—to handle multi-issue programs that can require hundreds or thousands of agent invocations.

How it works

Built on the AgentField runtime, SWE-AF orchestrates a control stack of planning, execution, and governance agents. It supports both single-repository and multi-repository modes, where it can coordinate changes across a primary application and its dependencies.

To maintain reliability, it uses three nested control loops:

• Inner loop: Handles single-issue retries based on QA/review feedback.
• Middle loop: Triggers an advisor to try new approaches or split work when the inner loop fails.
• Outer loop: Restructures the remaining task dependency graph (DAG) when escalated failures occur.

Who it’s for

It is designed for developers and organizations looking to automate complex software engineering tasks, from refactoring and hardening flows to adding new features across multiple codebases, without requiring manual human intervention for every step.

Highlights

• Factory Architecture: Uses a coordinated control stack rather than a single-agent wrapper.
• Multi-Model Support: Allows assigning different LLMs to different roles (e.g., Opus for coding, Haiku for QA).
• Multi-Repo Orchestration: Can manage coordinated changes across a primary app and shared libraries or microservices.
• Continual Learning: Can inject discovered conventions and failure patterns into downstream issues when enabled.
• Longe-run Reliability: Supports checkpointed execution to resume builds after interruptions.
• Agent-Scale Parallelism: Uses isolated git worktrees to allow large fan-out without branch collisions.