Enterprise Multi-Agent AI Operations Platform
Coordinate specialized AI agents, enterprise knowledge, tools, memory, and human approvals through one extensible platform.
English | Retained Chinese documentation
AgentForge AI is an enterprise-oriented foundation for multi-agent applications. It combines agent execution, retrieval-augmented generation (RAG), long-term memory, Model Context Protocol (MCP) tools, agent-to-agent (A2A) communication, and governance services behind a shared API gateway and operator interface.
This repository is a customized and extended derivative of atliliw/agent-platform. The upstream service architecture and implementation remain attributable to atliliw and contributors. AgentForge branding, the Business Agent Pack, catalog validation, integration code, tests, and supporting documentation are AgentForge-specific additions; no upstream endorsement is implied.
AgentForge AI is a software foundation, not a claim of production certification, regulatory compliance, customer adoption, or benchmark performance. Deployers remain responsible for security hardening, identity and access controls, tenant isolation, provider review, evaluation, observability, and approval policy.
- Multi-agent execution with streaming, tool calls, handoffs, checkpoints, interventions, and resumable sessions.
- RAG knowledge services with document ingestion, chunking, BM25 and vector retrieval, and Qdrant-backed indexing.
- Episodic, semantic, and working-memory services with recall and consolidation controls.
- Built-in and remote MCP tool discovery and execution.
- A2A agent discovery and task dispatch over gRPC and HTTP endpoints.
- Reusable agent skills with progressive loading.
- Harness services for guardrails, approvals, evaluations, prompts, workflows, SLOs, traces, session replay, and cost analytics.
- A tenant-aware Gin gateway, Go microservices, React operator interface, Docker Compose topologies, and OpenTelemetry collection.
- A five-role Business Agent Pack for governed operations, revenue, customer, risk, and executive-analysis workflows.
- Catalog validation and an adapter that integrates the business-agent definitions with the existing agent service.
- Evidence, uncertainty, and human-approval boundaries embedded in business-agent instructions.
- English architecture, configuration, deployment, API, development, and business-agent documentation.
- Lightweight pull-request CI for Go formatting, build, vet, tests, and frontend compilation.
flowchart TB
U[Users and API Clients] --> G[API Gateway]
G --> R[Agent and Multi-Agent Runtime]
R --> B[Business Agents]
R --> A[General AI Agents]
B --> O[Orchestration and Handoffs]
A --> O
O --> K[RAG Knowledge]
O --> M[Long-Term Memory]
O --> T[MCP Tools]
K --> X[A2A Communication]
M --> X
T --> X
X --> P[LLM and Tool Providers]
H[Governance and Human Approval] -. policies and checkpoints .-> R
H -. approval boundaries .-> B
H -. evaluation and observability .-> O
The gateway is the external HTTP boundary. Backend services use gRPC and retain separate responsibilities for chat, agents, knowledge, memory, A2A communication, tools, and governance. See Architecture for service topology, trust boundaries, failure modes, and approval flows.
Fresh installations add five AgentForge definitions alongside the existing upstream default agents:
| Agent | Purpose | Human boundary |
|---|---|---|
| Business Operations Director | Coordinate cross-functional analysis and handoffs | High-impact plans remain recommendations |
| Revenue Intelligence Agent | Analyze revenue operations and scenarios | No pricing, forecast, CRM, financial, or outbound execution |
| Customer Operations Agent | Synthesize cases and draft resolutions | No account changes, refunds, cancellations, or message sending |
| Risk and Compliance Review Agent | Review evidence against policy and risk criteria | No legal conclusion, filing, certification, or regulatory action |
| Executive Briefing Agent | Produce decision-ready, evidence-aware briefs | Briefs do not approve or execute decisions |
The pack uses verified built-in tools and explicit handoff contracts. Existing installations are not silently reseeded when their agent store is already populated. For the complete catalog contract, flows, tool policy, and extension guide, see Business Agents.
- Cross-functional operating reviews with specialist handoffs and explicit decision owners.
- Internal knowledge assistants using hybrid retrieval, contextual memory, and governed tools.
- Revenue and pipeline analysis based on traceable inputs and stated assumptions.
- Customer-operations case synthesis and draft resolution communications.
- Risk and compliance review that reserves legal, regulatory, privacy, and financial decisions for qualified humans.
- Executive briefings that preserve evidence, uncertainty, dissent, and approval requirements.
- Multi-agent workflows with checkpoints, evaluation, trace collection, and cost visibility.
| Layer | Technology |
|---|---|
| Backend | Go 1.22, Gin, gRPC, Protocol Buffers |
| Persistence and retrieval | MongoDB, SQLite, Qdrant, Redis |
| Agent integration | MCP tools, A2A services, OpenAI-compatible LLM endpoint support |
| Frontend | React 19, TypeScript, Ant Design 6, TanStack Query, Zustand, React Flow, Monaco, ECharts, Tailwind CSS 4, Vite |
| Observability | OpenTelemetry Collector |
| Deployment | Docker and Docker Compose |
The default example configuration targets DashScope/Qwen through an OpenAI-compatible endpoint. The Go module remains agent-platform for compatibility with the upstream codebase.
Prerequisites: Docker with Compose and a supported LLM API key.
# Generate the gitignored service configuration files.
bash scripts/init-config.sh sk-your-dashscope-key
# Windows PowerShell alternative:
pwsh scripts/init-config.ps1 sk-your-dashscope-key
# Build and start the full topology.
docker compose -f docker/docker-compose.yaml up -d --build
# Verify the gateway.
curl http://localhost:9000/healthOpen the frontend at http://localhost:8888 or call the gateway at http://localhost:9000. For a reduced local topology without the full browser, desktop, and observability sidecars, use:
docker compose -f docker/docker-compose.simple.yaml up -d --buildReal service configuration and credentials are gitignored. Review Configuration and Deployment before exposing any service.
.
├── .github/workflows/ # Pull-request and branch CI
├── configs/ # Shared configuration examples
├── docker/ # Full and reduced Compose topologies
├── docs/ # English and retained Chinese documentation
├── frontend/ # React operator interface
├── pkg/ # Shared Go packages, generated APIs, and business agents
├── proto/ # Protocol Buffer contracts
├── scripts/ # Configuration and development helpers
└── services/ # Gateway, agents, chat, knowledge, memory, A2A, MCP, and harness services
| Topic | Reference |
|---|---|
| Documentation index | docs/README.md |
| Architecture and trust boundaries | docs/en/architecture.md |
| Business Agent Pack | docs/en/business-agents.md |
| Configuration | docs/en/configuration.md |
| Deployment | docs/en/deployment.md |
| API reference | docs/en/api-reference.md |
| Development | docs/en/development.md |
| Upstream and fork attribution | NOTICE.md |
AgentForge AI preserves its upstream derivation, Git history, authorship, service identifiers, Go module path, and MIT license declaration. Upstream code is not presented as original AgentForge work. AgentForge-specific modifications are identified separately and do not imply sponsorship, partnership, certification, or endorsement by the upstream project.
See the Attribution Notice for the complete statement.
Licensed under the MIT License. The upstream copyright and permission notice remain intact. AgentForge-specific modifications are distributed under the same license; redistribution should preserve the license, attribution notice, and applicable authorship history.