How gagriglobal2025/forgeai works
Forge AI is a config-driven multi-agent workforce that turns a project brief into versioned SDLC artifacts. Its principal workflow loads the project and task DAG, dispatches dependency-ready agents, gathers their inputs and project context, calls an LLM, then stores and validates each output. Human approval, retries, rework, and escalation govern task completion. The README documents the domain stages and parallel branches; sampled source confirms the runtime and memory interactions. Some stage-level links are documented rather than directly visible in sampled implementation.
Main entry points: run.py Workflow entry · scheduler.py Task scheduler · base_agent.py Agent runtime · agents.yaml Product and architecture
Components 19
Run coordination
- Workflow entry
run.py - Task DAG
orchestrator/dag_engine.py - Task scheduler
orchestrator/scheduler.py
SDLC workforce
- Product and architecture
configs/agents.yaml - UX and accessibility
configs/agents.yaml - Frontend delivery
configs/agents.yaml - Backend delivery
configs/agents.yaml - QA and security
configs/agents.yaml - DevOps and handoff
configs/agents.yaml - Agent registry
agents/registry.py - Agent runtime
agents/base_agent.py
Shared memory
- Project context
memory/global_memory.py - Agent scratch
memory/agent_memory.py - Versioned artifacts
memory/artifact_store.py
LLM and events
- LLM client
llm/client.py - Event bus
orchestrator/event_bus.py
Other components
- Project operator
- Approver
- Anthropic or OpenAI
Connections 27
- Project operator → Workflow entry: starts run
README.md - Workflow entry → Task DAG: loads workflow
run.py - Workflow entry → Agent registry: loads agents
run.py - Workflow entry → Project context: seeds brief
run.py - Workflow entry → Task scheduler: configures and starts
run.py - Workflow entry → Agent scratch: creates
run.py - Workflow entry → Versioned artifacts: creates
run.py - Workflow entry → Event bus: creates and subscribes
run.py - Task scheduler → Task DAG: resolves ready tasks
orchestrator/scheduler.py - Task scheduler → Agent registry: dispatches agents
orchestrator/scheduler.py - Agent registry → Agent runtime: instantiates
agents/registry.py - Task scheduler → Event bus: emits task events
orchestrator/scheduler.py - Task scheduler → Approver: requests approval
orchestrator/scheduler.py - Approver → Task scheduler: approves or rejects
orchestrator/scheduler.py - Agent runtime → Project context: reads context
agents/base_agent.py - Agent runtime → Agent scratch: tracks attempts
agents/base_agent.py - Agent runtime → Versioned artifacts: reads inputs and writes output
agents/base_agent.py - Agent runtime → LLM client: requests completion
agents/base_agent.py - LLM client → Anthropic or OpenAI: calls selected API
llm/client.py - Agent runtime → Event bus: publishes outcomes
agents/base_agent.py - Task scheduler → Product and architecture: runs configured stages
README.md - Product and architecture → UX and accessibility: branches to design
README.md - UX and accessibility → Frontend delivery: feeds frontend
README.md - Product and architecture → Backend delivery: feeds backend
README.md - Frontend delivery → QA and security: enters QA
README.md - Backend delivery → QA and security: enters QA
README.md - QA and security → DevOps and handoff: advances to release
README.md