Intelligent Multi-Agent Orchestration for Autonomous Software Development

Based on Anthropic Managed Agents architecture. Weave orchestrates multiple LLM agents (planner, generator, evaluator) to automate the full software development lifecycle via LLM-driven dynamic DAG generation and execution.

DAG is the loom, Agent is the shuttle. Multiple agents collaborate by role, weaving requirements into complete software.

中文文档 | Architecture | Contributing | Changelog | Roadmap

• LLM-Driven DAG Orchestration -- Planner agent generates execution DAGs dynamically, adapting to failures in real-time
• Multi-Model Routing -- Assign different LLM models per agent role (e.g., Opus for planning, Sonnet for coding)
• Agent Memory -- Persistent cross-session memory with scope promotion (PRIVATE → SESSION → GLOBAL)
• Self-Learning -- Automatic pattern analysis from execution history, feeding optimization hints back to the planner
• Impact Analysis -- Pre-execution impact prediction and post-execution change verification
• DAG Templates -- Reusable YAML templates to skip LLM planning for recurring task patterns
• Skills System -- YAML-based prompt templates for single-agent invocations with variable substitution
• MCP Integration -- Model Context Protocol client for tool discovery and execution via stdio transport
• Web Console -- Real-time DAG monitoring, job management, and alert dashboard
• Approval Workflow -- Human-in-the-loop gate for high-risk operations
• Multiple Backends -- Local or git worktree isolation, with Docker sandbox support

• Python 3.11+
• An Anthropic API key (or OpenAI-compatible endpoint)

git clone https://github.com/yaogang1991/weave.git
cd weave
pip install -r requirements.txt

# Set your API key
export ANTHROPIC_API_KEY="sk-ant-..."

# One-command plan + execute
python main.py run "Build a REST API for todo items"

Or plan first, then execute:

python main.py plan "Build a REST API for user authentication"
python main.py execute ./data/plans/plan_xxx.json

# Plan and execute in one step
python main.py run "Add OAuth2 support to the API"

# With project-specific agents
python main.py run "Design login page" --project ./my-project --max-parallel 5

# Using a DAG template (skip LLM planning)
python main.py run "Build Todo API" --template build_api --var feature=Todo --var language=Python

# Terminal 1: Start worker
python main.py worker --concurrency 1

# Terminal 2: Submit task
python main.py submit "Build a REST API for user auth"

# Terminal 3: Monitor
python main.py list --status running
python main.py tickets --status pending

# Non-interactive mode
export WEAVE_NON_INTERACTIVE=true
python main.py worker --non-interactive

python main.py serve

python main.py viz
# Open http://localhost:8765 for the dashboard

┌──────────────────────────────────────────────────────────────┐
│                     Orchestrator Layer                        │
│   Planner  ·  Generator  ·  Evaluator                        │
├──────────────────────────────────────────────────────────────┤
│              Session Manager (Append-Only Event Log)          │
├──────────────────────────────────────────────────────────────┤
│                     Weave Core (Dumb Loop)                    │
│   Agent Worker  ←  Tool Registry  ←  Guardrails              │
├──────────────────────────────────────────────────────────────┤
│   Sandbox  ·  Git  ·  Reporter                                │
├──────────────────────────────────────────────────────────────┤
│   Memory  ·  Learning  ·  Impact Analysis                     │
└──────────────────────────────────────────────────────────────┘

Four-layer architecture: Orchestrator → Session Manager → Weave Core → Execution Layer

For the full architecture document, see ARCHITECTURE.md.

Create .weave/config.yaml in your project:

guardrails:
  permission_mode: default
  max_file_size: 100000

memory:
  enabled: true
  max_entries: 500

backend:
  type: local

See docs/config_reference.md for the full configuration reference.

Register project-specific agents in .weave/agents.yaml:

agents:
  - id: ui_designer
    name: UI Designer
    skills: [ui_design, react_component_dev, tailwind_css]
    constraints: [Only modifies frontend/src/]

The orchestrator discovers these automatically and assigns them during planning.

• Single-user scenario (no multi-tenancy)
• File-based storage (no external database required)
• Single-machine execution (no distributed mode)
• Impact analysis supports Python import resolution only

• Architecture -- Full system architecture and component details
• Contributing -- Development setup and PR process
• Changelog -- Release history
• Roadmap -- Milestone history and future plans
• Config Reference -- All configuration options
• Developer Guide -- Extending agents, tools, and backends
• Specs -- Per-module engineering specifications
• ADRs -- Architecture Decision Records

Apache License 2.0

Copyright 2026 yaogang1991