SWE-AF provides an autonomous software engineering runtime, enabling a full AI engineering team to handle tasks from planning to shipping code via a single API call. It targets developers and teams seeking to automate complex software development workflows, offering a scalable solution for production-grade PRs and end-to-end code delivery.

How It Works

SWE-AF operates as a coordinated "engineering factory" rather than a simple agent wrapper. It employs a control stack of planning, execution, and governance agents that adapt in real-time. Its "hardness-aware execution" intelligently scales effort based on task difficulty, triggering deeper adaptation and DAG-level replanning for complex issues. The architecture supports multi-repository orchestration, agent-scale parallelism via isolated worktrees, and fleet-scale orchestration through AgentField. Continual learning mechanisms inject discovered conventions and failure patterns into downstream tasks.

Quick Start & Requirements

• Primary Install: Local installation involves Python 3.12+, pip install -e ".[dev]". Docker and Railway one-click deployment options are also available.
• Prerequisites: Python 3.12+, AgentField control plane (af), AI provider API keys (Anthropic, OpenRouter, OpenAI, Google), and a GitHub Personal Access Token (PAT) with repo scope for draft PR creation.
• Links: LICENSE

Highlighted Details

• Autonomous team roles include Product Manager, Architect, Coders, Reviewers, and Testers.
• Supports multi-repository builds, coordinating changes across multiple codebases.
• Adaptive control loops (inner, middle, outer) manage task difficulty and replan the execution DAG.
• Achieved a 95/100 benchmark score with Claude Haiku and MiniMax M2.5, outperforming other models.
• Offers multi-model and multi-provider support, allowing model assignment per role.
• Features continual learning, explicit compromise tracking, and checkpointing for build resumption.

Maintenance & Community

SWE-AF is in public beta. Contribution guidelines, a code of conduct, and a security policy are available within the repository.

Licensing & Compatibility

The project is licensed under the Apache 2.0 license. No specific compatibility restrictions for commercial use are noted, though reliance on third-party LLM APIs implies associated costs and terms of service.

Limitations & Caveats

As a public beta, SWE-AF may have undiscovered issues or incomplete features. Its functionality is dependent on the availability and performance of external LLM APIs, requiring API keys and incurring usage costs.