Client loginBook a call
Build
Custom Software DevelopmentSystems built to your process
Web Application DevelopmentReact, Next.js, Rails, Django
Mobile App DevelopmentiOS, Android, React Native
App DevelopmentWeb and mobile apps, end to end
SaaS Product DevelopmentMulti-tenant from day one
MVP DevelopmentFirst release in 6–10 weeks
Enterprise SoftwareSSO, audit trails, scale
Connect & operate
API Development & IntegrationREST, GraphQL, webhooks
Cloud & DevOps EngineeringAWS, GCP, Azure, Terraform
QA & Software TestingAutomated suites, load, E2E
Legacy ModernizationStrangler-fig migrations
Cybersecurity EngineeringHardening and remediation
Data EngineeringPipelines, warehouses, reporting
Design & teams
UI/UX DesignInterfaces people finish tasks in
Prototyping & DiscoveryScope before you spend
Staff AugmentationSenior engineers on your team
Dedicated Dev TeamsA full pod, one contract
White-Label DevelopmentEngineering under your brand
Technology ArchitectureDecide once, correctly
Project ManagementDelivery you can audit
Start hereScope a build in one callBring a spec, a wireframe, or a paragraph. You leave with an architecture, a timeline, and a number.Book a scoping call
All software development services →How we price work →Talk to an engineer →
AI software
AI Application DevelopmentFull-stack AI products
AI Agent DevelopmentMulti-step autonomous agents
AI SaaS DevelopmentMulti-tenant AI products
Workflow AutomationReplace manual ops
Custom LLM AppsBranded internal assistants
LLM & data systems
Private LLM DevelopmentSelf-hosted Llama, Mistral, Qwen
RAG DevelopmentCited answers over your docs
MCP & WebMCPMake your systems agent-callable
Vector Database Setuppgvector, Pinecone, Qdrant
Document IntelligenceOCR and extraction pipelines
Vibe coding
AI-Assisted BuildsFast, with senior review
Rapid PrototypingClickable in days
Concept to CodeIdea to working software
Code Review & CleanupMake AI output shippable
Fractional AI TeamOngoing build capacity
Ready to ship?Put AI where the work isAgents, RAG, and private LLMs wired into the systems your team already uses — not a chatbot bolted to a homepage.Discuss an AI project
AI development services →AI readiness checklist →AI vs. traditional development →
Domain firstWe learn your workflow before we model itRegulated, operational, or high-volume — the constraints belong in the schema, not in a training doc.Talk about your domain
Selected workSoftware we own and operateMost agencies show screenshots. We run the products — the same stack, uptime, and on-call we hand to clients.View the portfolio
Plan smarterEstimate before you commitCost ranges, scope templates, and the questions we ask in discovery — free, no form.Open the cost calculator
Real conversationsTalk with a technical leadNo SDR, no discovery gauntlet. The person on the call is the one who scopes the build.Book a call
AI Coding Agents · Agent-Field

SWE-AF

SWE-AF is an autonomous multi-agent system that orchestrates software engineering tasks (planning, coding, testing, review) end-to-end via a single API call. It operates as a coordinated factory of specialized agents rather than a single coder loop, adapting strategy based on issue complexity and maintaining learned patterns across builds.

Talk to DEV.coDiscuss Implementation
Source: GitHub — github.com/Agent-Field/SWE-AF
900
GitHub stars
149
Forks
Python
Primary language
Apache-2.0
License (OSI-approved)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
RepositoryAgent-Field/SWE-AF
OwnerAgent-Field
Primary languagePython
LicenseApache-2.0 — OSI-approved
Stars900
Forks149
Open issues1
Latest releaseUnknown
Last updated2026-07-08
Sourcehttps://github.com/Agent-Field/SWE-AF

What SWE-AF is

SWE-AF is a Python-based agentic orchestrator built on AgentField that deploys role-specialized LLM agents (planner, architect, coder, reviewer, QA, merger) with multi-model support (Claude, OpenRouter, OpenAI, Google), dependency-aware DAG scheduling, isolated git worktrees for parallel execution, and checkpointed state for resumable builds. It claims 253.8x throughput improvement over subprocess-based execution via persistent interpreter pools.

Quickstart

Get the SWE-AF source

Clone the repository and explore it locally.

terminalbash
git clone https://github.com/Agent-Field/SWE-AF.gitcd SWE-AF# follow the project's README for install & configuration

Need it deployed, integrated, or customized instead? DEV.co ships production installs.

Best use cases

Multi-Issue Feature Development Across Multiple Repos

Coordinate simultaneous changes in primary applications and dependent libraries (shared SDKs, monorepo sub-projects, microservices). Multi-repo mode with role-based routing (primary/dependency) handles orchestration.

Autonomous PR Generation at Scale

Generate and ship production-grade PRs with full test coverage, code review, and merge automation. Demonstrated on real repo (PR #179: 10 issues, 217 tests passing, $19.23 total cost).

Continuous Autonomous Engineering Pipeline

Deploy fleet-scale orchestration via AgentField for thousands of concurrent agent invocations. Continual learning mode propagates conventions and failure patterns discovered early into downstream issues.

Implementation considerations

  • Requires Python 3.12+; AgentField runtime (version ≥0.1.87 for CLI); LLM API keys configured for chosen provider (Claude, OpenAI, Google, OpenRouter).
  • Multi-model configuration per role (planner, coder, reviewer, QA, etc.) allows cost/quality tuning; no default sensible configuration documented for typical use cases.
  • Checkpointed execution supports resumable builds after crashes, but checkpoint storage backend (local, remote, database) and recovery SLO not specified.
  • Isolated git worktrees enable parallel agent execution without branch collision; requires sufficient disk I/O and git operational overhead for large concurrent fleets.
  • Enable learning mode (`enable_learning=true`) to inject conventions discovered early into downstream issues, but no explicit API to audit or rollback learned patterns.

When to avoid it — and what to weigh

  • Strict Determinism or Audit-Trail Requirements — Multi-agent orchestration with LLM-driven adaptation introduces non-deterministic behavior. No clear versioning of agent decision trees or replay guarantees for compliance-critical workflows.
  • Closed-Source or Proprietary Code Constraints — System requires sending code to external LLM providers (Claude, OpenAI, Google, OpenRouter). No mention of on-premise or air-gapped deployment options.
  • Minimal LLM Inference Budget — Demonstrated example cost $19.23 for 10 issues with 79 agent invocations. Scale and cost grow with issue complexity; no cost optimization boundaries documented.
  • Legacy Monolithic Codebases Without Test Coverage — System relies on autonomous test execution and verification. Low test coverage or highly coupled legacy code will limit adaptation capability and increase failure rates.

License & commercial use

Licensed under Apache License 2.0 (Apache-2.0), a permissive OSI-approved license permitting commercial use, modification, and redistribution with liability/warranty disclaimers and trademark restrictions.

Apache 2.0 permits commercial use. However, system depends on third-party LLM APIs (Claude, OpenAI, etc.) whose terms govern actual commercial deployment. Verify LLM provider ToS for your use case. No indemnification, SLA, or support terms are provided by SWE-AF; requires review of external dependencies.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityModerate
DEV.co fitGood
Assessment confidenceMedium
Security considerations

System sends application code to external LLM providers (Claude, OpenAI, Google, OpenRouter). Verify provider data retention and compliance policies for your codebase. No mention of code sanitization, secret masking, or audit logging. git operations assume SSH/HTTPS auth; no discussion of credential rotation or key management. No security audit or threat model published.

Alternatives to consider

GitHub Copilot Workspace / Copilot PR

Microsoft-native PR automation via copilot; integrated GitHub experience. Narrower scope (single-agent coder loop) and closed-source; does not support multi-repo orchestration or custom model assignment.

Aider (aider.chat)

Open-source, multi-model CLI agent for code generation and editing. Lighter-weight than SWE-AF; supports chat-driven iteration. No autonomous multi-agent orchestration, planning, or factory governance.

Continue.dev IDE plugin

IDE-native AI coding assistant with multi-model support. Interactive loop; requires human validation. No autonomous PR generation or end-to-end engineering workflow orchestration.

Software development agency

Build on SWE-AF with DEV.co software developers

Start with SWE-AF's quick-start guide. Requires Python 3.12+, AgentField runtime, and LLM provider API key. Deploy locally first, then scale to multi-node fleet orchestration.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

SWE-agentopen-sweagent-of-empires

Related on DEV.co

Explore the category and the services that help you build with it.

AI Coding AgentsAI Development ServicesAI-Powered Application DevelopmentCustom Software DevelopmentAPI Development ServicesTalk to DEV.co

SWE-AF FAQ

Can I run SWE-AF on my own infrastructure without sending code to external LLMs?
Not clearly stated. System supports multiple LLM providers via OpenRouter, but no on-premise or self-hosted LLM integration is documented. Requires review of provider options.
How much will it cost to autonomously build a typical feature?
Example PR #179 (10 issues, haiku-class models): $19.23. Cost scales with issue complexity, agent invocations, and chosen models. No cost estimation API or budget controls documented.
What happens if an agent fails mid-build?
Checkpointed execution supports resumable builds, but checkpoint storage, recovery SLO, and retry limits are not specified. Requires review of checkpoint mechanism and failure modes.
Does SWE-AF work with private GitHub repos?
Likely (GitHub PR integration shown), but authentication mechanism (SSH keys, tokens, OAuth) and permission model not documented. Requires review of setup guide.

Senior engineers for your next build

Our engineers ship ai coding agents software for a living. DEV.co provides software development services, web development services, and ongoing support for teams standardizing on tools like SWE-AF.

Ready to Automate Your Engineering Workflow?

Start with SWE-AF's quick-start guide. Requires Python 3.12+, AgentField runtime, and LLM provider API key. Deploy locally first, then scale to multi-node fleet orchestration.

Talk to DEV.coSchedule a Consultation