Agent-Field/SWE-AF

What it solves

SWE-AF is an autonomous software engineering runtime that transforms a single API call into a complete engineering team's output. It solves the problem of scaling AI coding from simple, single-loop scripts to complex, multi-issue programs that require coordination between product management, architecture, coding, review, and testing.

How it works

Built on the AgentField framework, the system operates as a coordinated "factory" of agents rather than a single coder. It employs three nested control loops to handle tasks of varying difficulty:

• Inner loop: Handles single-issue retries based on QA or review failures.
• Middle loop: Triggers an "issue advisor" to try new approaches or split work when the inner loop is exhausted.
• Outer loop: Uses a "replanner" to restructure the entire remaining task dependency graph (DAG) when failures escalate.

It supports both single-repository and multi-repository modes, allowing it to coordinate changes across a primary application and its dependencies. It can use different LLMs for different roles (e.g., Opus for coding, Haiku for QA) and supports multiple runtimes including Claude Code, OpenCode, and Codex.

Who it’s for

Software engineers and organizations looking to automate complex feature development, refactoring, and bug fixing across one or more repositories without manual human intervention.

Highlights

• Multi-role coordination: Orchestrates PMs, architects, coders, reviewers, and testers.
• Hardness-aware execution: Adapts the planning depth and strategy based on the difficulty of the issue.
• Multi-repo support: Coordinates changes across primary apps and dependency libraries.
• Continual learning: Injects discovered conventions and failure patterns into downstream tasks.
• Agent-scale parallelism: Uses isolated git worktrees to allow large-scale fan-out without branch collisions.
• Checkpointing: Supports resume_build to recover from crashes or interruptions.