SWE-AF

SWE-AF is an Autonomous Engineering Team Runtime built on the AgentField control plane. Users submit a natural language goal and repository URL via a single API call, and the system orchestrates 22 specialized Agents (Product Manager, Architect, Coder, Reviewer, Tester, etc.) to produce a GitHub Draft PR with architecture summary and tech debt inventory. A typical build orchestrates 400–500+ Agent calls through a six-stage pipeline: Plan + Git Init → Execute Issue DAG → Verify-Fix Loop → Repo Finalize → Push + Draft PR.

Core Capabilities

• One-call engineering factory: from goal to Draft PR in a single API call
• Single-repo and multi-repo modes: cross-codebase coordination (app + shared libs/microservices)
• Three-layer adaptive control: inner loop (per-issue retry, max 5) → mid loop (Issue Advisor strategy) → outer loop (Replanner DAG-level replanning)
• Difficulty-aware execution: simple issues pass fast, complex issues trigger deeper adaptation
• Runtime plan mutation: Replanner dynamically adds/removes issues and dependencies
• Graceful degradation + explicit debt tracking: debt is classified, graded, and propagated downstream

Concurrency & Isolation

• Structured concurrency + barrier synchronization: DAG-level parallelism with gate sequence
• Agent isolation + semantic merging: independent git worktree per issue, Merger Agent for semantic merge
• Cross-Agent knowledge propagation: early findings spread to downstream issues when learning enabled
• Persistent checkpoint resumption: DAGState serialized to checkpoint.json at stage boundaries

Multi-Model Scheduling

• Dual runtimes: claude_code (Claude backend) and open_code (OpenRouter / OpenAI / Google / Anthropic)
• Per-role model assignment: e.g., coder→opus, qa→haiku, default→sonnet across 16 role keys
• Risk-proportional resource allocation: needs_deeper_qa routes issues to 2-call or 4-call paths

Governance & Traceability

• Each Agent has a DID (Decentralized Identifier) derived via BIP-44
• Every Agent call generates a VC (Verifiable Credential) with caller/target DID, I/O hashes, timestamp, cryptographic signature
• All VCs form a workflow chain for end-to-end traceability

22 Specialized Agents

• Planning: Product Manager, Architect, Tech Lead, Sprint Planner, Issue Writer
• Execution: Coder, QA, Code Reviewer, QA Synthesizer, Retry Advisor, Issue Advisor, Replanner
• Git/Merge: Git Init, Workspace Setup, Merger, Integration Tester, Workspace Cleanup
• Verification/Finalization: Verifier, Fix Generator, Repo Finalizer, GitHub PR Creator

Deployment Options: Railway one-click (SWE-AF + AgentField + PostgreSQL), Docker Compose (with worker scaling), local install (Python >= 3.12).

Use Cases: Autonomous feature development, cross-repo coordinated changes, large-scale refactoring, API-triggered CI automation. Sibling project SEC-AF enables automated security auditing after SWE-AF builds.

Unconfirmed: Benchmark 95/100 score only in README without independent third-party verification; JPMorgan, ServiceNow, NIH use AgentField platform but SWE-AF usage unconfirmed; swe-fast mode differences undocumented; initial release date unknown.