Multi-Agent Software Engineering Team

Autonomous issue-to-PR orchestration for AI software engineering workflows.

Why This Exists

Single-agent coding loops collapse when one context has to hold issue context, repo architecture, multiple implementation branches, review feedback, test logs, and cost accounting. This project splits the workflow into explicit roles:

• architect: converts a GitHub issue into a subtask DAG
• coders: claim subtasks and work in isolated worktrees/sandboxes
• merge coordinator: stages branches and records conflicts
• reviewer: reviews the merged diff only
• tester: verifies in a clean environment and stores artifacts
• reporter: exports cost, trace, benchmark, and handoff-failure summaries

Current Status

This repository is a production-oriented scaffold. It includes the core contracts, adapters, runners, CLI commands, tests, and deployment skeleton, but it is not yet a published benchmark result. Live autonomous runs require real provider credentials and a sandbox backend.

Features

• GitHub issue ingestion through gh
• Append-only JSONL task board with worker leases
• Config-driven provider factory for Anthropic, OpenAI, Gemini, and local fake providers
• Local, Docker, and Daytona runner interfaces
• Worktree lifecycle manager
• Scope guard for coder patches
• Staging merge coordinator and conflict resolver hook
• Reviewer and tester model adapters
• Artifact store, status summaries, preflight diagnostics, and schema export
• Cost accounting, tracing fallback, benchmark report structures
• Kubernetes starter manifests and Dockerfile

Install

python3 -m venv .venv
. .venv/bin/activate
python3 -m pip install -e .
scripts/self_check.py

Quick Start

Run local diagnostics:

mast preflight --repo .

Export the message schema:

mast schema

Run a local dry issue intake flow:

mast run --issue ramshetty01/multiagent-software-team#97 --run-id demo --board /tmp/mast-board.jsonl --artifact-dir /tmp/mast-artifacts
mast status --board /tmp/mast-board.jsonl --run-id demo

Configuration

Local mode uses fake providers. Production mode requires provider credentials:

cp .env.example .env

Required production variables:

• GITHUB_TOKEN
• ANTHROPIC_API_KEY
• OPENAI_API_KEY
• GOOGLE_API_KEY
• LANGFUSE_PUBLIC_KEY
• LANGFUSE_SECRET_KEY

See docs/configuration.md.

Architecture

flowchart LR
  Issue[GitHub Issue] --> Intake
  Intake --> Board[(Task Board)]
  Board --> Architect
  Architect --> Coders
  Coders --> Merge
  Merge --> Reviewer
  Reviewer -->|feedback| Coders
  Reviewer --> Tester
  Tester -->|pass| PR[Pull Request]
  Tester -->|fail| Coders
  Board --> Report[Reports and Traces]

Loading

Detailed docs:

• Architecture
• Run lifecycle
• Operator setup
• Benchmarks and evaluation
• Operations runbook
• Troubleshooting

Evaluation

The repository includes evaluation structures and a placeholder 50-issue file. Real SWE-bench Pro scoring requires licensed issue IDs and live model execution. The required protocol and artifact contract are documented in docs/benchmarks.md.

Tracked metrics:

• pass@1
• single-agent baseline pass@1
• wall-clock speedup
• token and dollar cost by role
• reviewer false-approval rate
• handoff-failure histogram

Security

This project executes untrusted repository code. Treat all target repositories as hostile unless explicitly allowlisted. Use Docker or Daytona isolation for production runs. See docs/threat-model.md.

Limitations

• No published SWE-bench Pro score yet.
• Live model tests are not run in default CI.
• Kubernetes manifests are starter manifests, not a validated production cluster.
• The default JSONL task board is intended for local and small-scale runs.
• Provider prompts and schemas are still evolving.

Roadmap

• Publish a real E2E demo run.
• Replace placeholder benchmark issue IDs with licensed SWE-bench Pro IDs.
• Add full LangGraph checkpoint persistence.
• Add live-provider integration test suite behind explicit environment gates.
• Add docs site and benchmark report.

License

License is tracked in LICENSE.